{ "source_file": "0930", "source_platform": "知乎专栏", "author": "stone", "title_listed": "图像语义分割综述", "keywords": "语义分割,综述", "domain": "图像处理", "content_type": "综述", "is_paid_flag": "否", "url": "https://zhuanlan.zhihu.com/p/37801090", "publish_date": "2018", "corpus_stage": "应用", "notes": "图像语义分割领域综述", "id": "dd6eafdfdf38", "collected_at": "2026-10-02T18:44:07", "title": "图像语义分割综述 - 知乎", "body_markdown": "[![](https://zhuanlan.zhihu.com/p/37801090)](javascript:void\\(0\\))\n![ZhiHu logo](../images/body-e1fa91dbdb48b8055184.png)\n[](https://www.zhihu.com)\n[关注](https://www.zhihu.com/signin?next=%2Ffollow)[推荐](https://www.zhihu.com/signin?next=%2F)[热榜](https://www.zhihu.com/signin?next=%2Fhot)[专栏](https://www.zhihu.com/signin?next=%2Fcolumn-square)[圈子](https://www.zhihu.com/signin?next=%2Fring-feeds)[AI WorksBeta ](https://www.zhihu.com/project-square)[故事](https://www.zhihu.com/fiore/h5/vip-web)\n​\n[直答](https://zhida.zhihu.com/)\n切换模式\n登录/注册\n图像语义分割综述\n首发于[图像语义分割](https://www.zhihu.com/column/c_197474183)\n切换模式\n![图像语义分割综述](../images/000.img)\n# 图像语义分割综述\n[![stone](../images/035.jpg)](https://www.zhihu.com/people/big-big-stone)\n[stone](https://www.zhihu.com/people/big-big-stone)\n收藏之前点个赞呗\n[收录于 · 图像语义分割](https://www.zhihu.com/column/c_197474183)\n666 人赞同了该文章\n## 更多语义分割文章，请看专栏：[图像语义分割](https://zhuanlan.zhihu.com/c_197474183)\n## 一、什么是语义分割\n语义分割是在像素级别上的分类，属于同一类的像素都要被归为一类，因此语义分割是从像素级别来理解图像的。比如说如下的照片，属于人的像素都要分成一类，属于摩托车的像素也要分成一类，除此之外还有背景像素也被分为一类。注意语义分割不同于实例分割，举例来说，如果一张照片中有多个人，对于语义分割来说，只要将所由人的像素都归为一类，但是[实例分割](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=2&q=%E5%AE%9E%E4%BE%8B%E5%88%86%E5%89%B2&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLlrp7kvovliIblibIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjIsInpkX3Rva2VuIjpudWxsfQ.mIqTxwKvPGgMJhI_qss2hxllVMDt4_I972vZZqRRNdY&zhida_source=entity)还要将不同人的像素归为不同的类。也就是说实例分割比语义分割更进一步。\n![](../images/004.jpg)\n图1(a). 原图\n![](../images/005.jpg)\n图1(b). 分割结果\n## 二、语义分割的思路\n## I.传统方法\n在深度学习方法流行之前，[TextonForest](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=TextonForest&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJUZXh0b25Gb3Jlc3QiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.tDo578lT6QU7lVi0wv9qJ33qa28huJvWjuU79WIYW8E&zhida_source=entity)和基于[随机森林分类器](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E9%9A%8F%E6%9C%BA%E6%A3%AE%E6%9E%97%E5%88%86%E7%B1%BB%E5%99%A8&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLpmo_mnLrmo67mnpfliIbnsbvlmagiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.4sE8OeiriXQKVDeMq9FxsoIc4D7sySBkIJTeK_cRZ5Q&zhida_source=entity)等语义分割方法是用得比较多的方法。不过在[深度卷积网络](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E6%B7%B1%E5%BA%A6%E5%8D%B7%E7%A7%AF%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLmt7Hluqbljbfnp6_nvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.pN71qi0gSeHekdxepuyDeTXCtwXKgxfUfGesaL4ZmKU&zhida_source=entity)流行之后，深度学习方法比传统方法提升了很多，所以这里就不详细讲传统方法了。\n## II.深度学习方法\n深度学习方法在语义分割上得到了巨大成功，深度学习方法解决语义分割问题可以概括为几种思路。下面进行详细介绍。\n**1.[Patch classification](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=Patch+classification&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJQYXRjaCBjbGFzc2lmaWNhdGlvbiIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjczMTE2ODgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.6j6_O7-VNCbegZz2LwjC0QsanHf8jAh2FFuDLbqgKLE&zhida_source=entity)**\n最初的深度学习方法应用于图像分割就是Patch classification。Patch classification方法，顾名思义，图像是切成块喂给深度模型的，然后对像素进行分类。使用图像块的主要原因是因为全连接层需要固定大小的图像。\n**2.[全卷积](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E5%85%A8%E5%8D%B7%E7%A7%AF&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLlhajljbfnp68iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.MUiFa7KC-oEkpz2EwomPb-7fNzwDV1DxUqEht43CVj0&zhida_source=entity)方法**\n2014年，[全卷积网络](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E5%85%A8%E5%8D%B7%E7%A7%AF%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLlhajljbfnp6_nvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.Pu3v-WhX_d0wDW2iqDi_oE3iu6-16NtKrx5ag51oP0I&zhida_source=entity)（FCN）横空出世，FCN将网络全连接层用卷积取代，因此使任意图像大小的输入都变成可能，而且速度比Patch classification方法快很多。\n尽管移除了[全连接层](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=3&q=%E5%85%A8%E8%BF%9E%E6%8E%A5%E5%B1%82&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLlhajov57mjqXlsYIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjMsInpkX3Rva2VuIjpudWxsfQ.80k_1cvXxQ5X0hH_i3hFAdYC-57rATFat4VsDeSASKc&zhida_source=entity)，但是CNN模型用于语义分割还存在一个问题，就是[下采样](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E4%B8%8B%E9%87%87%E6%A0%B7&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLkuIvph4fmoLciLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.Q39eS23ZZmLrhCnP1VH58UifhVdV5Nt8bBIh-Uv7ZS0&zhida_source=entity)操作（比如，pooling）。pooling操作可以扩大[感受野](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E6%84%9F%E5%8F%97%E9%87%8E&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLmhJ_lj5fph44iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.y6k34AfmRDUN3bx0jdh6Q3AV_09_vTeDAK1o06Gz9T8&zhida_source=entity)因而能够很好地整合上下文信息（context中文称为语境或者上下文，通俗的理解就是综合了更多的信息来进行决策），对high-level的任务（比如分类），这是很有效的。但同时，由于pooling下采样操作，使得分辨率降低，因此削弱了位置信息，而语义分割中需要score map和原图对齐，因此需要丰富的位置信息。\n**3.encoder-decoder架构**\nencoder-decoder是基于FCN的架构。encoder由于pooling逐渐减少空间维度，而decoder逐渐恢复空间维度和细节信息。通常从encoder到decoder还有shortcut connetction（捷径连接，也就是[跨层连接](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E8%B7%A8%E5%B1%82%E8%BF%9E%E6%8E%A5&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLot6jlsYLov57mjqUiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.feh1yJ7hENgVnUD8hDy_MwIcjdDG8VZrnRT5Wm8rkJw&zhida_source=entity)）。其中[U-net](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=U-net&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJVLW5ldCIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjczMTE2ODgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.nTzxew-VSjNPM4oEpDazwkR0NzdIJUU_Qry-XKB3pZo&zhida_source=entity)就是这种架构很流行的一种，如下图：\n![](../images/006.jpg)\n**4.空洞卷积**\ndilated/atrous （空洞卷积）架构，这种结构代替了pooling，一方面它可以保持空间分辨率，另外一方面它由于可以扩大感受野因而可以很好地整合上下文信息。如下图：\n![](../images/016.jpg)\n**5.条件随机场**\n除了以上思路，还有一种对分割结果进行后处理的方法，那就是[条件随机场](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=2&q=%E6%9D%A1%E4%BB%B6%E9%9A%8F%E6%9C%BA%E5%9C%BA&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLmnaHku7bpmo_mnLrlnLoiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjIsInpkX3Rva2VuIjpudWxsfQ.Iou0Ntg0O-a0Hu2Zr_g__XGyd7AcOLO9KTnirfEH1kU&zhida_source=entity)(Conditional Random Fields ([CRF](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=CRF&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJDUkYiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.3O98pcysBjBoP9UMPZhCQgn0u-eHaPAOp8u-QRoSPfo&zhida_source=entity)s))后处理用来改善分割效果。[DeepLab系列](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=DeepLab%E7%B3%BB%E5%88%97&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJEZWVwTGFi57O75YiXIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6NzMxMTY4OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.1S5ETZH1xMKw5FjWG7bgaB4IQ3ggUg_I2gOHXsMIpVo&zhida_source=entity)文章基本都采用这种后处理方法，可以较好地改善分割结果，如下图：\n![](../images/018.jpg)\n## 三、深度学习语义分割方法\n## 方法总览\n现在的深度学习语义分割模型基本上都是基于FCN发展而来的，它是开山鼻祖，一张图概括FCN的延伸方法：\n![](../images/009.jpg)\n## 各方法的详细信息\n![](../images/010.jpg)\n## 各方法简要介绍\n下面简单总结一些从FCN进行改进的几种架构，关于每种架构的详细解读请看专栏中其他文章。\n 1. [FCN](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23fcn)\n 2. [SegNet](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23segnet)\n 3. [Dilated Convolutions](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23dilation)\n 4. [DeepLab (v1 & v2)](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23deeplab)\n 5. [RefineNet](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23refinenet)\n 6. [PSPNet](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23pspnet)\n 7. [Large Kernel Matters](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23large-kernel)\n 8. [DeepLab v3](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23deeplabv3)\n\n\n## 1.FCN\n**论文信息**\n> Fully Convolutional Networks for Semantic Segmentation \n> Submitted on 14 Nov 2014 \n> [Arxiv Link](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/1411.4038)\n**主要贡献**\n 1. 使端对端的卷积语义分割网络变得流行起来。\n 2. 通过deconvolutional layers进行上采样。\n 3. 通过skip connection改善了上采样的粗糙度。\n\n\n**概要**\n 1. **全卷积化(Fully Convolutional)** ：用于解决逐像素([pixel](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=pixel&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJwaXhlbCIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjczMTE2ODgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.Tzyo1BVfycZAN6G2TXzGlnONVYlzNgB7liwFCNtkZvA&zhida_source=entity)-wise)的预测问题。通过将基础网络(例如VGG)最后面几个全连接层换成卷积层，可实现任意大小的图像输入，并且输出图像大小与输入相对应；\n 2. **[反卷积](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E5%8F%8D%E5%8D%B7%E7%A7%AF&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLlj43ljbfnp68iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.0rw3ZYgsZ6kmT2_dAhqU-So0X3ifdOw1i0HSHdXvkvM&zhida_source=entity)(deconvolution)** ：上采样操作，用于恢复图片尺寸，方便后续进行逐像素预测;\n 3. **跳跃结构(skip[architecture](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=architecture&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJhcmNoaXRlY3R1cmUiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.7gWRUVFoR9lAGj-m7dEe0yAPQOhNz71Wn3156HErDVQ&zhida_source=entity))**：用于融合高低层特征信息。通过跨层连接的结构，结合了网络浅层的细(fine-grain)粒度信息信息以及深层的粗糙(coarse)信息，以实现精准的分割任务。\n\n\n![](../images/011.jpg)\n**Benchmarks (VOC2012)**\n![](../images/012.jpg)\n评论：\nFCN是基于深度学习的语义分割的开山之作，尽管现在很多方法都超越了FCN，但它的思想仍然有很重要的意义。\n## 2. Segnet\n**论文信息**\n> SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation \n> Submitted on 2 Nov 2015 \n> [Arxiv Link](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/1511.00561)\n**主要贡献**\n * 使用Maxpooling indices来增强位置信息。\n\n\n**简要概述**\nFCN的upconvolution层+shortcut connections产生的分割图比较粗糙，因此SegNet增加了更多的[shortcut connections](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=2&q=shortcut+connections&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJzaG9ydGN1dCBjb25uZWN0aW9ucyIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjczMTE2ODgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MiwiemRfdG9rZW4iOm51bGx9.uM6L7zoQuoPEUL3asRvuxtRzhsSEIKLoW0pXPPbKtDo&zhida_source=entity)。不过，SegNet并不是直接将encoder的特征进行直接复制，而是对maxpooling中的indices进行复制，这使得SegNet的效率更高。\n![](../images/013.jpg)\nmaxpooling 的indices复制原理如下：\n![](../images/014.jpg)\n \n\n**Benchmarks (VOC2012)**\n![](../images/015.jpg)\n**评论**\n * FCN和SegNet都是encoder-decoder架构。\n * SegNet的benchmark表现太差了，不建议用这个网络。\n\n\n## 3. Dilated convolution\n**论文信息**\n> Multi-Scale Context Aggregation by Dilated Convolutions \n> Submitted on 23 Nov 2015 \n> [Arxiv Link](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/1511.07122)\n**创新点**\n 1. 使用[空洞卷积](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=3&q=%E7%A9%BA%E6%B4%9E%E5%8D%B7%E7%A7%AF&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLnqbrmtJ7ljbfnp68iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjMsInpkX3Rva2VuIjpudWxsfQ.MjWocpHkycUbaufokYQ-luJELpphPHft0tbrwP7LISk&zhida_source=entity)用来进行稠密预测（dense prediction）。\n 2. 提出上下文模块（context module），使用空洞卷积（Dilated Convolutions）来进行多尺度信息的的整合。\n\n\n**简要解释**\npooling操作可以增大感受野，对于图像分类任务来说这有很大好处，但由于pooling操作降低了分辨率，这对语义分割来说很不利。因此作者提出一种叫做dilated convolution的操作来解决这个问题。dilated卷积(在deeplab中称为atrous卷积)。可以很好地提升感受野的同时可以保持空间分辨率。\n![](../images/016.jpg)\n网络架构有两种，一种是前端网络，另外一种是前端网络+上下文模块，分别介绍如下：\n * 将VGG网络的最后两个pooling层给拿掉了，之后的[卷积层](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=2&q=%E5%8D%B7%E7%A7%AF%E5%B1%82&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLljbfnp6_lsYIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjIsInpkX3Rva2VuIjpudWxsfQ.W9cFhtU-T5Wdraw7hQ2d8EEpuqZr8thHLLIc1fJmFQ8&zhida_source=entity)被dilated 卷积取代。并且在pool3和pool4之间空洞卷积的空洞率=2，pool4之后的空洞卷积的[空洞率](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=2&q=%E7%A9%BA%E6%B4%9E%E7%8E%87&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLnqbrmtJ7njociLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjIsInpkX3Rva2VuIjpudWxsfQ.46jbWG-KfQv9qKqfLX74Zam_ZtnW27MrmCli8li_-gQ&zhida_source=entity)=4。作者将这种架构称为前端（front-end）。\n * 除了前端网络之外，作者还设计了一种叫做上下文模块（context module）的架构，加在前端网络之后。上下文木块中[级联](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E7%BA%A7%E8%81%94&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLnuqfogZQiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.ibS_tX7g1GwQ3DMZ08OCNYimacIy5pm1OQ_W7DDiAII&zhida_source=entity)了多种不同空洞率的空洞卷积，使得多尺度的上下文信息可以得到整合，从而改善前端网络预测的效果。**需要注意的是[前端网络](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=6&q=%E5%89%8D%E7%AB%AF%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLliY3nq6_nvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjYsInpkX3Rva2VuIjpudWxsfQ.J96gf4KUAUzF1i_EhvIif10J03JP6_xQJBOiSnVV7zs&zhida_source=entity)和上下文木块是分开训练的，因为作者在实验中发现，如果是联合在一起进行端对端的训练并不能改善性能。**\n\n\n**Benchmarks (VOC2012)**\n![](../images/017.jpg)\n**评论**\n需要特别注意的是，网络输出的分割图并不是和原始图像大小一样的，而是其1/8，需要对输出的分割图进行[线性插值](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E7%BA%BF%E6%80%A7%E6%8F%92%E5%80%BC&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLnur_mgKfmj5LlgLwiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.0FQndfaB6wyBnCdkmqP5Y9bRevAM6LeUXy5SAAXN3Ik&zhida_source=entity)才能得到最终的分割结果。这种做法也是很多其他的方法都使用的。\n## 4. DeepLab(v1,v2)\n**论文信息**\n> **v1** : Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs \n> Submitted on 22 Dec 2014 \n> [Arxiv Link](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/1412.7062) \n> **v2** : DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs \n> Submitted on 2 Jun 2016 \n> [Arxiv Link](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/1606.00915)\n**主要贡献**\n 1. 使用atrous卷积，也就是后来的空洞卷积，扩大感受野，保持分辨率。\n 2. 提出了atrous spatial [pyramid](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=pyramid&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJweXJhbWlkIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6NzMxMTY4OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.HkJZfhFZnTTstPIAtcZko2mDP82XgbePhnl-doXOIf4&zhida_source=entity) pooling (ASPP)，整合多尺度信息。\n 3. 使用全连接条件随机场（fully connected CRF)进行后处理，改善分割结果。\n\n\n**简要概述**\n 1. 空洞卷积可以在不增加参数的情况下增加感受野。\n 2. 通过两种方式来进行多尺度的处理：A.将原始图像的多种尺度喂给网络进行训练。B.通过平行的不同空洞率的[空洞卷积层](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E7%A9%BA%E6%B4%9E%E5%8D%B7%E7%A7%AF%E5%B1%82&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLnqbrmtJ7ljbfnp6_lsYIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.9G9yj4k7kmw15af_WQsFmGNPDMkmkJyfqd9lELyNs5c&zhida_source=entity)来获得。\n 3. 通过全连接条件随机场来进行后处理，以改善分割结果。\n\n\n![](../images/018.jpg)\n \n\n**Benchmarks (VOC2012)**\n![](../images/019.jpg)\n## 5. RefineNet\n**论文信息**\n> RefineNet: Multi-Path Refinement Networks for High-Resolution Semantic Segmentation \n> Submitted on 20 Nov 2016 \n> [Arxiv Link](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/1611.06612)\n**主要贡献**\n 1. 精心设计了encoder-decoder架构中的decoder部分，使得性能提升。\n 2. 整个网络的设计都遵循[residual](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=residual&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJyZXNpZHVhbCIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjczMTE2ODgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.qz3TNrii00Juk8hNMprDsoM5FrdEuuLqq7mheUzpAsc&zhida_source=entity) connections，网络表达能力更强，[梯度](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E6%A2%AF%E5%BA%A6&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLmoq_luqYiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.TbF6pT3h6eTMYPwvzMxzVCpbpyPgquulkT_xr1CudJk&zhida_source=entity)更容易[反向传播](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E5%8F%8D%E5%90%91%E4%BC%A0%E6%92%AD&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLlj43lkJHkvKDmkq0iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.7MzFQMWhsdtWnmIru1OFLU9KmY4EX97BQW1lK8npzws&zhida_source=entity)。\n\n\n**简要概述**\n作者提出空洞卷积方法应用于语义分割也是有缺点的，包括：\n * 因为使用了大分辨率的[feature map](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=feature+map&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJmZWF0dXJlIG1hcCIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjczMTE2ODgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.gRP8AFwXtjL-zUjp1dgUKdc4hlYt4efUpqPr85Kjvv0&zhida_source=entity)，因此计算代价大，并且需要大量的内存。对于这个问题，DeepLab的做法是只预测原始输入的1／8。\n\n\n本文提出使用encoder-decoder架构。encoder部分是RESNET-101。decoder具有RefineNet blocks，它将此前的RefineNet blocks的低分辨率特征和encoder部分[高分辨率](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E9%AB%98%E5%88%86%E8%BE%A8%E7%8E%87&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLpq5jliIbovqjnjociLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.I1XSsl6T9y0JdzgaVmrUkDhBmMjTcLI7uJ4ef9S4I_U&zhida_source=entity)特征进行concatenate/fuse。\n![](../images/020.jpg)\n![](../images/021.jpg)\n**Benchmarks (VOC2012)**\n![](../images/022.jpg)\n## 6. PSPNet\n**论文信息**\n> [Pyramid](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=Pyramid&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJQeXJhbWlkIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6NzMxMTY4OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.8cHFzuIcN9MZdbcJW3dzPG3-Y0VDEl3K_g7tgDg0nN0&zhida_source=entity) Scene Parsing Network \n> Submitted on 4 Dec 2016 \n> [Arxiv Link](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/1612.01105)\n**主要贡献**\n 1. 使用pyramid pooling整合context。\n 2. 使用auxiliary loss。\n\n\n**概要**\n[骨架网络](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E9%AA%A8%E6%9E%B6%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLpqqjmnrbnvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.Q-GB5AwmKw_jwzfIkDXBH6YjLmoom_Sq5wnlSepbV1M&zhida_source=entity)使用Resnet，并在此基础上加上pyramid pooling module。该模块用到了很多kernel大小不一的pooling 。将pooling的结果再上采样，经过concatenate进行融合。\n![](../images/023.jpg)\n在RESNET的第四阶段（即输入到金字塔池模块）之后，应用auxiliary loss。这种方法在别的地方也被称为intermediate supervision。\n**Benchmarks (VOC2012)**\n![](../images/024.jpg)\n## 以下两篇论文还没细看，看完了之后更新。\n## 7. Large Kernel Matters\npass\n## 8. deeplab v3\npass\n \n\n**参考**\n这篇文章主要是参考以下两篇文章的一个总结，加入了一些自己的理解，想看原文的直接打开：\n 1. [A 2017 Guide to Semantic Segmentation with Deep Learning](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23dilation)\n 2. [A Review on Deep Learning Techniques Applied to Semantic Segmentation](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/1704.06857)\n\n\n编辑于 2018-06-10 00:30\n[深度学习（Deep Learning）](https://www.zhihu.com/topic/19813032)\n[机器学习](https://www.zhihu.com/topic/19559450)\n[计算机视觉](https://www.zhihu.com/topic/19590195)\n​赞同 666​​19 条评论​1183 ​74 \n​分享\n​申请转载​\n​\n![](../images/body-7e064b35ba4350074989.jpeg)\n未登录用户\n \n\n19 条评论\n默认\n最新\n[![dadream](../images/025.jpg)](https://www.zhihu.com/people/008813f5db257ab17c34c20f970a2a43)\n[dadream](https://www.zhihu.com/people/008813f5db257ab17c34c20f970a2a43)\n没有unet？\n2018-09-20\n​回复​3\n展开其他 1 条回复​\n[![tan知yi](../images/026.jpg)](https://www.zhihu.com/people/93e7e9bfdea1eafb71b842e2e45cd275)\n[tan知yi](https://www.zhihu.com/people/93e7e9bfdea1eafb71b842e2e45cd275)\n想知道传统的随机森林是如何做语义分割的![\\[谢邀\\]](https://pic2.zhimg.com/v2-6fe2283baa639ae1d7c024487f1d68c7.png)![\\[捂脸\\]](https://pic1.zhimg.com/v2-b62e608e405aeb33cd52830218f561ea.png)\n2020-12-28\n​回复​1\n[![寒青](../images/027.jpg)](https://www.zhihu.com/people/b2863314f4f336526ae871e34f21dfc7)\n[寒青](https://www.zhihu.com/people/b2863314f4f336526ae871e34f21dfc7)\n想问下，分割图不同类别不同颜色是如何实现的\n2020-07-04\n​回复​1\n展开其他 1 条回复​\n[![wong或许对了](../images/028.jpg)](https://www.zhihu.com/people/956c37aa6e474f888ba4b689b8cd9e43)\n[wong或许对了](https://www.zhihu.com/people/956c37aa6e474f888ba4b689b8cd9e43)\n大佬不讲讲OCRnet吗\n2023-10-13\n​回复​喜欢\n[![大山](../images/029.jpg)](https://www.zhihu.com/people/a4304b214b7d2909d7dd0b69aa90a791)\n[大山](https://www.zhihu.com/people/a4304b214b7d2909d7dd0b69aa90a791)\n很棒的综述，非常谢谢！\n2022-10-18\n​回复​喜欢\n[![知乎用户dwCPx8](../images/030.jpg)](https://www.zhihu.com/people/7bf35d9cf5bb04ea23b22572ea6c40a1)\n[知乎用户dwCPx8](https://www.zhihu.com/people/7bf35d9cf5bb04ea23b22572ea6c40a1)\n1\n2022-09-22\n​回复​喜欢\n[![Eric](../images/031.jpg)](https://www.zhihu.com/people/3d0cac71d63596fd1d20c0223954020c)\n[Eric](https://www.zhihu.com/people/3d0cac71d63596fd1d20c0223954020c)\n![\\[赞同\\]](https://pic2.zhimg.com/v2-419a1a3ed02b7cfadc20af558aabc897.png)\n2021-06-02\n​回复​喜欢\n[![月色血风暴](../images/032.jpg)](https://www.zhihu.com/people/a18eaeb489cced34ef81da46c4cb694f)\n[月色血风暴](https://www.zhihu.com/people/a18eaeb489cced34ef81da46c4cb694f)\n感谢~写得很棒\n2020-02-21\n​回复​喜欢\n[![storm](../images/033.jpg)](https://www.zhihu.com/people/81cfaab7a1b8b5524328bc14f5feaf3b)\n[storm](https://www.zhihu.com/people/81cfaab7a1b8b5524328bc14f5feaf3b)\n感谢！可以更新吗\n2019-11-02\n​回复​喜欢\n[![李文鑫](../images/034.jpg)](https://www.zhihu.com/people/1d9439567b386e7df3e19d60d1bc1c33)\n[李文鑫](https://www.zhihu.com/people/1d9439567b386e7df3e19d60d1bc1c33)\n但是上次看论文 没看懂的 看了这篇之后 会更懂 谢谢作者了\n2019-09-25\n​回复​喜欢\n点击查看全部评论\n![](../images/body-7e064b35ba4350074989.jpeg)\n未登录用户\n \n\n关于作者\n[![stone](../images/035.jpg)](https://www.zhihu.com/people/big-big-stone)\n[stone](https://www.zhihu.com/people/big-big-stone)\n收藏之前点个赞呗\n[ 回答 **30** ](https://www.zhihu.com/people/big-big-stone/answers)[ 文章 **15** ](https://www.zhihu.com/people/big-big-stone/posts)[ 关注者 **10,013** ](https://www.zhihu.com/people/big-big-stone/followers)\n​关注他​发私信\n大家都在搜\n换一换\n[江歌妈妈发长文尘埃终将落定704 万](https://www.zhihu.com/search?q=%E6%B1%9F%E6%AD%8C%E5%A6%88%E5%A6%88%E5%8F%91%E9%95%BF%E6%96%87%E5%B0%98%E5%9F%83%E7%BB%88%E5%B0%86%E8%90%BD%E5%AE%9A&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)热\n[中国男足 0-5 巴勒斯坦562 万](https://www.zhihu.com/search?q=%E4%B8%AD%E5%9B%BD%E7%94%B7%E8%B6%B3+0-5+%E5%B7%B4%E5%8B%92%E6%96%AF%E5%9D%A6&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)热\n[原央视主持人阿丘回应被通报419 万](https://www.zhihu.com/search?q=%E5%8E%9F%E5%A4%AE%E8%A7%86%E4%B8%BB%E6%8C%81%E4%BA%BA%E9%98%BF%E4%B8%98%E5%9B%9E%E5%BA%94%E8%A2%AB%E9%80%9A%E6%8A%A5&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)热\n[华为与赛力斯达成新五年合作409 万](https://www.zhihu.com/search?q=%E5%8D%8E%E4%B8%BA%E4%B8%8E%E8%B5%9B%E5%8A%9B%E6%96%AF%E8%BE%BE%E6%88%90%E6%96%B0%E4%BA%94%E5%B9%B4%E5%90%88%E4%BD%9C&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)热\n[迪拜航空确认航班发生事故398 万](https://www.zhihu.com/search?q=%E8%BF%AA%E6%8B%9C%E8%88%AA%E7%A9%BA%E7%A1%AE%E8%AE%A4%E8%88%AA%E7%8F%AD%E5%8F%91%E7%94%9F%E4%BA%8B%E6%95%85&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)热\n[东航回应网传空姐跪地道歉326 万](https://www.zhihu.com/search?q=%E4%B8%9C%E8%88%AA%E5%9B%9E%E5%BA%94%E7%BD%91%E4%BC%A0%E7%A9%BA%E5%A7%90%E8%B7%AA%E5%9C%B0%E9%81%93%E6%AD%89&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)\n[王楚钦：不知道为什么就是感觉累320 万](https://www.zhihu.com/search?q=%E7%8E%8B%E6%A5%9A%E9%92%A6%EF%BC%9A%E4%B8%8D%E7%9F%A5%E9%81%93%E4%B8%BA%E4%BB%80%E4%B9%88%E5%B0%B1%E6%98%AF%E6%84%9F%E8%A7%89%E7%B4%AF&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)热\n[江苏高考作文《衬衫的价格为 9 磅 15 便士》爆火268 万](https://www.zhihu.com/search?q=%E6%B1%9F%E8%8B%8F%E9%AB%98%E8%80%83%E4%BD%9C%E6%96%87%E3%80%8A%E8%A1%AC%E8%A1%AB%E7%9A%84%E4%BB%B7%E6%A0%BC%E4%B8%BA+9+%E7%A3%85+15+%E4%BE%BF%E5%A3%AB%E3%80%8B%E7%88%86%E7%81%AB&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)\n[女装网店开始用防拆带了265 万](https://www.zhihu.com/search?q=%E5%A5%B3%E8%A3%85%E7%BD%91%E5%BA%97%E5%BC%80%E5%A7%8B%E7%94%A8%E9%98%B2%E6%8B%86%E5%B8%A6%E4%BA%86&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)\n[孕妇骑车别车被司机踹翻257 万](https://www.zhihu.com/search?q=%E5%AD%95%E5%A6%87%E9%AA%91%E8%BD%A6%E5%88%AB%E8%BD%A6%E8%A2%AB%E5%8F%B8%E6%9C%BA%E8%B8%B9%E7%BF%BB&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)\n### 推荐阅读\n# [浅谈图像的语义分割 基于深度学习的图像语义分割综述 摘要 近些年来计算机视觉与机器学习研究者对图像语义分割问题越来越感兴趣，越来越多的应用场景需要精确且高效的分割技术，如自动驾驶、室内导航、甚至虚拟… Laayt ](https://zhuanlan.zhihu.com/p/100140440)[ ![图片语义分割深度学习算法要点回顾](../images/037.jpg)图片语义分割深度学习算法要点回顾 AI研习社 ](https://zhuanlan.zhihu.com/p/53233072)[ ![图像语义分割\\(20\\) 通过图像合成方法检测训练中未出现的类别未知的物体](../images/039.jpg)图像语义分割(20) 通过图像合成方法检测训练中未出现的类别未知的物体 赵磊发表于计算机科学... ](https://zhuanlan.zhihu.com/p/78193038)[ ![【图解AI】什么是语义分割、实例分割、全景分割](../images/body-0f434b1eb5c9d0554840.jpg)【图解AI】什么是语义分割、实例分割、全景分割 Jack Wang ](https://zhuanlan.zhihu.com/p/368904941) \n\n_想来知乎工作？请发送邮件到 jobs@zhihu.com_\n打开知乎App\n在「我的页」右上角打开扫一扫\n![](../images/body-0f9788d006ab0f4da762.png)\n其他扫码方式：微信\n下载知乎App\n无障碍模式\n验证码登录\n密码登录\n[开通机构号](https://zhuanlan.zhihu.com/org/signup)\n中国 +86\n获取短信验证码\n忘记密码\n登录/注册\n其他方式登录\n未注册手机验证后自动登录，注册即代表同意[《知乎协议》](https://www.zhihu.com/term/zhihu-terms)[《隐私保护指引》](https://www.zhihu.com/term/privacy)\n扫码下载知乎 App\n关闭二维码\n", "metadata": { "title": "图像语义分割综述 - 知乎", "description": "更多语义分割文章，请看专栏：图像语义分割一、什么是语义分割语义分割是在像素级别上的分类，属于同一类的像素都要被归为一类，因此语义分割是从像素级别来理解图像的。比如说如下的照片，属于人的像素都要分成一…", "keywords": "深度学习（Deep Learning）,机器学习,计算机视觉", "author": null, "og:title": "图像语义分割综述", "og:url": "https://zhuanlan.zhihu.com/p/37801090", "og:description": "更多语义分割文章，请看专栏：图像语义分割一、什么是语义分割语义分割是在像素级别上的分类，属于同一类的像素都要被归为一类，因此语义分割是从像素级别来理解图像的。比如说如下的照片，属于人的像素都要分成一…", "og:image": "https://picx.zhimg.com/v2-464813a7c07f9e03c9d3db1ddc632f14_720w.jpg?source=172ae18b", "og:type": "article", "og:site_name": "知乎专栏" }, "paywall": { "listed_paid": false, "detected": false, "marker": "", "public_only": false, "policy": "付费/会员内容只保留页面公开可见部分，不绕过付费墙" }, "crawl": { "success": true, "status_code": 200, "error": "" }, "code_blocks": [], "images": [ { "status": "http_403", "url": "https://zhuanlan.zhihu.com/p/37801090", "src": "https://zhuanlan.zhihu.com/p/37801090", "alt": "", "downloaded": false, "discovery": "saved_markdown", "retried_at": "2026-10-20T18:05:35Z" } ], "audio": [], "videos": [], "links": [ { "type": "internal", "href": "https://www.zhihu.com" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2Ffollow" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2F" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2Fhot" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2Fcolumn-square" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2Fring-feeds" }, { "type": "internal", "href": "https://www.zhihu.com/project-square" }, { "type": "internal", "href": "https://www.zhihu.com/fiore/h5/vip-web" }, { "type": "internal", "href": "https://zhida.zhihu.com/" }, { "type": "internal", "href": "https://www.zhihu.com/column/c_197474183" }, { "type": "internal", "href": "https://www.zhihu.com/people/big-big-stone" }, { "type": "internal", "href": "https://zhuanlan.zhihu.com/c_197474183" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=2&q=%E5%AE%9E%E4%BE%8B%E5%88%86%E5%89%B2&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLlrp7kvovliIblibIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjIsInpkX3Rva2VuIjpudWxsfQ.mIqTxwKvPGgMJhI_qss2hxllVMDt4_I972vZZqRRNdY&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=TextonForest&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJUZXh0b25Gb3Jlc3QiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.tDo578lT6QU7lVi0wv9qJ33qa28huJvWjuU79WIYW8E&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E9%9A%8F%E6%9C%BA%E6%A3%AE%E6%9E%97%E5%88%86%E7%B1%BB%E5%99%A8&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLpmo_mnLrmo67mnpfliIbnsbvlmagiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.4sE8OeiriXQKVDeMq9FxsoIc4D7sySBkIJTeK_cRZ5Q&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E6%B7%B1%E5%BA%A6%E5%8D%B7%E7%A7%AF%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLmt7Hluqbljbfnp6_nvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.pN71qi0gSeHekdxepuyDeTXCtwXKgxfUfGesaL4ZmKU&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=Patch+classification&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJQYXRjaCBjbGFzc2lmaWNhdGlvbiIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjczMTE2ODgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.6j6_O7-VNCbegZz2LwjC0QsanHf8jAh2FFuDLbqgKLE&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E5%85%A8%E5%8D%B7%E7%A7%AF&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLlhajljbfnp68iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.MUiFa7KC-oEkpz2EwomPb-7fNzwDV1DxUqEht43CVj0&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E5%85%A8%E5%8D%B7%E7%A7%AF%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLlhajljbfnp6_nvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.Pu3v-WhX_d0wDW2iqDi_oE3iu6-16NtKrx5ag51oP0I&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=3&q=%E5%85%A8%E8%BF%9E%E6%8E%A5%E5%B1%82&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLlhajov57mjqXlsYIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjMsInpkX3Rva2VuIjpudWxsfQ.80k_1cvXxQ5X0hH_i3hFAdYC-57rATFat4VsDeSASKc&zhida_source=entity" }, { "type": "external", "href": "javascript:void(0)" } ], "repair_version": 2, "repaired_at": "2026-10-20T17:32:33Z", "source_url": "https://zhuanlan.zhihu.com/p/37801090", "delivery_status": "unverified", "discovery_status": "http_403", "missing_reason": "源页面未重新核验:http_403；已按保留正文补采图片", "repair_error": "", "actual_video_files": [], "expected_saved_body_images": 1, "image_discovery": "saved_markdown", "source_body_markdown": "[![](https://zhuanlan.zhihu.com/p/37801090)](javascript:void\\(0\\))\n![ZhiHu logo](https://static.zhihu.com/heifetz/assets/wechat-share-logo.39ea9ecd.png)\n[](https://www.zhihu.com)\n[关注](https://www.zhihu.com/signin?next=%2Ffollow)[推荐](https://www.zhihu.com/signin?next=%2F)[热榜](https://www.zhihu.com/signin?next=%2Fhot)[专栏](https://www.zhihu.com/signin?next=%2Fcolumn-square)[圈子](https://www.zhihu.com/signin?next=%2Fring-feeds)[AI WorksBeta ](https://www.zhihu.com/project-square)[故事](https://www.zhihu.com/fiore/h5/vip-web)\n​\n[直答](https://zhida.zhihu.com/)\n切换模式\n登录/注册\n图像语义分割综述\n首发于[图像语义分割](https://www.zhihu.com/column/c_197474183)\n切换模式\n![图像语义分割综述](https://picx.zhimg.com/70/v2-464813a7c07f9e03c9d3db1ddc632f14_1440w.image?source=172ae18b&biz_tag=Post)\n# 图像语义分割综述\n[![stone](https://picx.zhimg.com/36f310418_l.jpg?source=32738c0c&needBackground=1)](https://www.zhihu.com/people/big-big-stone)\n[stone](https://www.zhihu.com/people/big-big-stone)\n收藏之前点个赞呗\n[收录于 · 图像语义分割](https://www.zhihu.com/column/c_197474183)\n666 人赞同了该文章\n## 更多语义分割文章，请看专栏：[图像语义分割](https://zhuanlan.zhihu.com/c_197474183)\n## 一、什么是语义分割\n语义分割是在像素级别上的分类，属于同一类的像素都要被归为一类，因此语义分割是从像素级别来理解图像的。比如说如下的照片，属于人的像素都要分成一类，属于摩托车的像素也要分成一类，除此之外还有背景像素也被分为一类。注意语义分割不同于实例分割，举例来说，如果一张照片中有多个人，对于语义分割来说，只要将所由人的像素都归为一类，但是[实例分割](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=2&q=%E5%AE%9E%E4%BE%8B%E5%88%86%E5%89%B2&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLlrp7kvovliIblibIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjIsInpkX3Rva2VuIjpudWxsfQ.mIqTxwKvPGgMJhI_qss2hxllVMDt4_I972vZZqRRNdY&zhida_source=entity)还要将不同人的像素归为不同的类。也就是说实例分割比语义分割更进一步。\n![](https://pic4.zhimg.com/v2-43df0837c26b53347ffbb03454812f87_1440w.jpg)\n图1(a). 原图\n![](https://pic1.zhimg.com/v2-5ff5f87333bf4a6702e05a00373e079c_1440w.jpg)\n图1(b). 分割结果\n## 二、语义分割的思路\n## I.传统方法\n在深度学习方法流行之前，[TextonForest](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=TextonForest&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJUZXh0b25Gb3Jlc3QiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.tDo578lT6QU7lVi0wv9qJ33qa28huJvWjuU79WIYW8E&zhida_source=entity)和基于[随机森林分类器](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E9%9A%8F%E6%9C%BA%E6%A3%AE%E6%9E%97%E5%88%86%E7%B1%BB%E5%99%A8&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLpmo_mnLrmo67mnpfliIbnsbvlmagiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.4sE8OeiriXQKVDeMq9FxsoIc4D7sySBkIJTeK_cRZ5Q&zhida_source=entity)等语义分割方法是用得比较多的方法。不过在[深度卷积网络](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E6%B7%B1%E5%BA%A6%E5%8D%B7%E7%A7%AF%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLmt7Hluqbljbfnp6_nvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.pN71qi0gSeHekdxepuyDeTXCtwXKgxfUfGesaL4ZmKU&zhida_source=entity)流行之后，深度学习方法比传统方法提升了很多，所以这里就不详细讲传统方法了。\n## II.深度学习方法\n深度学习方法在语义分割上得到了巨大成功，深度学习方法解决语义分割问题可以概括为几种思路。下面进行详细介绍。\n**1.[Patch classification](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=Patch+classification&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJQYXRjaCBjbGFzc2lmaWNhdGlvbiIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjczMTE2ODgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.6j6_O7-VNCbegZz2LwjC0QsanHf8jAh2FFuDLbqgKLE&zhida_source=entity)**\n最初的深度学习方法应用于图像分割就是Patch classification。Patch classification方法，顾名思义，图像是切成块喂给深度模型的，然后对像素进行分类。使用图像块的主要原因是因为全连接层需要固定大小的图像。\n**2.[全卷积](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E5%85%A8%E5%8D%B7%E7%A7%AF&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLlhajljbfnp68iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.MUiFa7KC-oEkpz2EwomPb-7fNzwDV1DxUqEht43CVj0&zhida_source=entity)方法**\n2014年，[全卷积网络](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E5%85%A8%E5%8D%B7%E7%A7%AF%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLlhajljbfnp6_nvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.Pu3v-WhX_d0wDW2iqDi_oE3iu6-16NtKrx5ag51oP0I&zhida_source=entity)（FCN）横空出世，FCN将网络全连接层用卷积取代，因此使任意图像大小的输入都变成可能，而且速度比Patch classification方法快很多。\n尽管移除了[全连接层](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=3&q=%E5%85%A8%E8%BF%9E%E6%8E%A5%E5%B1%82&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLlhajov57mjqXlsYIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjMsInpkX3Rva2VuIjpudWxsfQ.80k_1cvXxQ5X0hH_i3hFAdYC-57rATFat4VsDeSASKc&zhida_source=entity)，但是CNN模型用于语义分割还存在一个问题，就是[下采样](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E4%B8%8B%E9%87%87%E6%A0%B7&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLkuIvph4fmoLciLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.Q39eS23ZZmLrhCnP1VH58UifhVdV5Nt8bBIh-Uv7ZS0&zhida_source=entity)操作（比如，pooling）。pooling操作可以扩大[感受野](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E6%84%9F%E5%8F%97%E9%87%8E&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLmhJ_lj5fph44iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.y6k34AfmRDUN3bx0jdh6Q3AV_09_vTeDAK1o06Gz9T8&zhida_source=entity)因而能够很好地整合上下文信息（context中文称为语境或者上下文，通俗的理解就是综合了更多的信息来进行决策），对high-level的任务（比如分类），这是很有效的。但同时，由于pooling下采样操作，使得分辨率降低，因此削弱了位置信息，而语义分割中需要score map和原图对齐，因此需要丰富的位置信息。\n**3.encoder-decoder架构**\nencoder-decoder是基于FCN的架构。encoder由于pooling逐渐减少空间维度，而decoder逐渐恢复空间维度和细节信息。通常从encoder到decoder还有shortcut connetction（捷径连接，也就是[跨层连接](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E8%B7%A8%E5%B1%82%E8%BF%9E%E6%8E%A5&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLot6jlsYLov57mjqUiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.feh1yJ7hENgVnUD8hDy_MwIcjdDG8VZrnRT5Wm8rkJw&zhida_source=entity)）。其中[U-net](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=U-net&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJVLW5ldCIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjczMTE2ODgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.nTzxew-VSjNPM4oEpDazwkR0NzdIJUU_Qry-XKB3pZo&zhida_source=entity)就是这种架构很流行的一种，如下图：\n![](https://pica.zhimg.com/v2-6ef05b0d8349ff806c83790f7462c30e_1440w.jpg)\n**4.空洞卷积**\ndilated/atrous （空洞卷积）架构，这种结构代替了pooling，一方面它可以保持空间分辨率，另外一方面它由于可以扩大感受野因而可以很好地整合上下文信息。如下图：\n![](https://pic3.zhimg.com/v2-b1ff163f7a014186d69fdc9cdf74f10c_1440w.jpg)\n**5.条件随机场**\n除了以上思路，还有一种对分割结果进行后处理的方法，那就是[条件随机场](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=2&q=%E6%9D%A1%E4%BB%B6%E9%9A%8F%E6%9C%BA%E5%9C%BA&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLmnaHku7bpmo_mnLrlnLoiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjIsInpkX3Rva2VuIjpudWxsfQ.Iou0Ntg0O-a0Hu2Zr_g__XGyd7AcOLO9KTnirfEH1kU&zhida_source=entity)(Conditional Random Fields ([CRF](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=CRF&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJDUkYiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.3O98pcysBjBoP9UMPZhCQgn0u-eHaPAOp8u-QRoSPfo&zhida_source=entity)s))后处理用来改善分割效果。[DeepLab系列](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=DeepLab%E7%B3%BB%E5%88%97&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJEZWVwTGFi57O75YiXIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6NzMxMTY4OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.1S5ETZH1xMKw5FjWG7bgaB4IQ3ggUg_I2gOHXsMIpVo&zhida_source=entity)文章基本都采用这种后处理方法，可以较好地改善分割结果，如下图：\n![](https://pica.zhimg.com/v2-e4e863acf8aff610b89598f3e8f7d28a_1440w.jpg)\n## 三、深度学习语义分割方法\n## 方法总览\n现在的深度学习语义分割模型基本上都是基于FCN发展而来的，它是开山鼻祖，一张图概括FCN的延伸方法：\n![](https://picx.zhimg.com/v2-94cf04ce72e41e7dc3c5495e287b4fd1_1440w.jpg)\n## 各方法的详细信息\n![](https://pic1.zhimg.com/v2-84b0d01fd9de2d5ba0db880928127a8c_1440w.jpg)\n## 各方法简要介绍\n下面简单总结一些从FCN进行改进的几种架构，关于每种架构的详细解读请看专栏中其他文章。\n 1. [FCN](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23fcn)\n 2. [SegNet](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23segnet)\n 3. [Dilated Convolutions](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23dilation)\n 4. [DeepLab (v1 & v2)](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23deeplab)\n 5. [RefineNet](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23refinenet)\n 6. [PSPNet](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23pspnet)\n 7. [Large Kernel Matters](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23large-kernel)\n 8. [DeepLab v3](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23deeplabv3)\n\n\n## 1.FCN\n**论文信息**\n> Fully Convolutional Networks for Semantic Segmentation \n> Submitted on 14 Nov 2014 \n> [Arxiv Link](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/1411.4038)\n**主要贡献**\n 1. 使端对端的卷积语义分割网络变得流行起来。\n 2. 通过deconvolutional layers进行上采样。\n 3. 通过skip connection改善了上采样的粗糙度。\n\n\n**概要**\n 1. **全卷积化(Fully Convolutional)** ：用于解决逐像素([pixel](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=pixel&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJwaXhlbCIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjczMTE2ODgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.Tzyo1BVfycZAN6G2TXzGlnONVYlzNgB7liwFCNtkZvA&zhida_source=entity)-wise)的预测问题。通过将基础网络(例如VGG)最后面几个全连接层换成卷积层，可实现任意大小的图像输入，并且输出图像大小与输入相对应；\n 2. **[反卷积](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E5%8F%8D%E5%8D%B7%E7%A7%AF&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLlj43ljbfnp68iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.0rw3ZYgsZ6kmT2_dAhqU-So0X3ifdOw1i0HSHdXvkvM&zhida_source=entity)(deconvolution)** ：上采样操作，用于恢复图片尺寸，方便后续进行逐像素预测;\n 3. **跳跃结构(skip[architecture](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=architecture&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJhcmNoaXRlY3R1cmUiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.7gWRUVFoR9lAGj-m7dEe0yAPQOhNz71Wn3156HErDVQ&zhida_source=entity))**：用于融合高低层特征信息。通过跨层连接的结构，结合了网络浅层的细(fine-grain)粒度信息信息以及深层的粗糙(coarse)信息，以实现精准的分割任务。\n\n\n![](https://pic2.zhimg.com/v2-ea95c8eafcd5c2a6ed0ca74585131f8d_1440w.jpg)\n**Benchmarks (VOC2012)**\n![](https://pic4.zhimg.com/v2-695ca771e3ec237d279c93462f66f6e1_1440w.jpg)\n评论：\nFCN是基于深度学习的语义分割的开山之作，尽管现在很多方法都超越了FCN，但它的思想仍然有很重要的意义。\n## 2. Segnet\n**论文信息**\n> SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation \n> Submitted on 2 Nov 2015 \n> [Arxiv Link](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/1511.00561)\n**主要贡献**\n * 使用Maxpooling indices来增强位置信息。\n\n\n**简要概述**\nFCN的upconvolution层+shortcut connections产生的分割图比较粗糙，因此SegNet增加了更多的[shortcut connections](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=2&q=shortcut+connections&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJzaG9ydGN1dCBjb25uZWN0aW9ucyIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjczMTE2ODgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MiwiemRfdG9rZW4iOm51bGx9.uM6L7zoQuoPEUL3asRvuxtRzhsSEIKLoW0pXPPbKtDo&zhida_source=entity)。不过，SegNet并不是直接将encoder的特征进行直接复制，而是对maxpooling中的indices进行复制，这使得SegNet的效率更高。\n![](https://pic2.zhimg.com/v2-542d601001e471c8afc2cbbb124d75a5_1440w.jpg)\nmaxpooling 的indices复制原理如下：\n![](https://pic2.zhimg.com/v2-6eaee89e8473580911f96886005e34af_1440w.jpg)\n \n\n**Benchmarks (VOC2012)**\n![](https://pic1.zhimg.com/v2-df10b8850427277ef6fbd249d4d841da_1440w.jpg)\n**评论**\n * FCN和SegNet都是encoder-decoder架构。\n * SegNet的benchmark表现太差了，不建议用这个网络。\n\n\n## 3. Dilated convolution\n**论文信息**\n> Multi-Scale Context Aggregation by Dilated Convolutions \n> Submitted on 23 Nov 2015 \n> [Arxiv Link](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/1511.07122)\n**创新点**\n 1. 使用[空洞卷积](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=3&q=%E7%A9%BA%E6%B4%9E%E5%8D%B7%E7%A7%AF&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLnqbrmtJ7ljbfnp68iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjMsInpkX3Rva2VuIjpudWxsfQ.MjWocpHkycUbaufokYQ-luJELpphPHft0tbrwP7LISk&zhida_source=entity)用来进行稠密预测（dense prediction）。\n 2. 提出上下文模块（context module），使用空洞卷积（Dilated Convolutions）来进行多尺度信息的的整合。\n\n\n**简要解释**\npooling操作可以增大感受野，对于图像分类任务来说这有很大好处，但由于pooling操作降低了分辨率，这对语义分割来说很不利。因此作者提出一种叫做dilated convolution的操作来解决这个问题。dilated卷积(在deeplab中称为atrous卷积)。可以很好地提升感受野的同时可以保持空间分辨率。\n![](https://pic3.zhimg.com/v2-b1ff163f7a014186d69fdc9cdf74f10c_1440w.jpg)\n网络架构有两种，一种是前端网络，另外一种是前端网络+上下文模块，分别介绍如下：\n * 将VGG网络的最后两个pooling层给拿掉了，之后的[卷积层](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=2&q=%E5%8D%B7%E7%A7%AF%E5%B1%82&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLljbfnp6_lsYIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjIsInpkX3Rva2VuIjpudWxsfQ.W9cFhtU-T5Wdraw7hQ2d8EEpuqZr8thHLLIc1fJmFQ8&zhida_source=entity)被dilated 卷积取代。并且在pool3和pool4之间空洞卷积的空洞率=2，pool4之后的空洞卷积的[空洞率](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=2&q=%E7%A9%BA%E6%B4%9E%E7%8E%87&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLnqbrmtJ7njociLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjIsInpkX3Rva2VuIjpudWxsfQ.46jbWG-KfQv9qKqfLX74Zam_ZtnW27MrmCli8li_-gQ&zhida_source=entity)=4。作者将这种架构称为前端（front-end）。\n * 除了前端网络之外，作者还设计了一种叫做上下文模块（context module）的架构，加在前端网络之后。上下文木块中[级联](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E7%BA%A7%E8%81%94&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLnuqfogZQiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.ibS_tX7g1GwQ3DMZ08OCNYimacIy5pm1OQ_W7DDiAII&zhida_source=entity)了多种不同空洞率的空洞卷积，使得多尺度的上下文信息可以得到整合，从而改善前端网络预测的效果。**需要注意的是[前端网络](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=6&q=%E5%89%8D%E7%AB%AF%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLliY3nq6_nvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjYsInpkX3Rva2VuIjpudWxsfQ.J96gf4KUAUzF1i_EhvIif10J03JP6_xQJBOiSnVV7zs&zhida_source=entity)和上下文木块是分开训练的，因为作者在实验中发现，如果是联合在一起进行端对端的训练并不能改善性能。**\n\n\n**Benchmarks (VOC2012)**\n![](https://pic2.zhimg.com/v2-2012959fdea2997a3c5fa0c99e0eb099_1440w.jpg)\n**评论**\n需要特别注意的是，网络输出的分割图并不是和原始图像大小一样的，而是其1/8，需要对输出的分割图进行[线性插值](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E7%BA%BF%E6%80%A7%E6%8F%92%E5%80%BC&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLnur_mgKfmj5LlgLwiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.0FQndfaB6wyBnCdkmqP5Y9bRevAM6LeUXy5SAAXN3Ik&zhida_source=entity)才能得到最终的分割结果。这种做法也是很多其他的方法都使用的。\n## 4. DeepLab(v1,v2)\n**论文信息**\n> **v1** : Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs \n> Submitted on 22 Dec 2014 \n> [Arxiv Link](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/1412.7062) \n> **v2** : DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs \n> Submitted on 2 Jun 2016 \n> [Arxiv Link](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/1606.00915)\n**主要贡献**\n 1. 使用atrous卷积，也就是后来的空洞卷积，扩大感受野，保持分辨率。\n 2. 提出了atrous spatial [pyramid](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=pyramid&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJweXJhbWlkIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6NzMxMTY4OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.HkJZfhFZnTTstPIAtcZko2mDP82XgbePhnl-doXOIf4&zhida_source=entity) pooling (ASPP)，整合多尺度信息。\n 3. 使用全连接条件随机场（fully connected CRF)进行后处理，改善分割结果。\n\n\n**简要概述**\n 1. 空洞卷积可以在不增加参数的情况下增加感受野。\n 2. 通过两种方式来进行多尺度的处理：A.将原始图像的多种尺度喂给网络进行训练。B.通过平行的不同空洞率的[空洞卷积层](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E7%A9%BA%E6%B4%9E%E5%8D%B7%E7%A7%AF%E5%B1%82&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLnqbrmtJ7ljbfnp6_lsYIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.9G9yj4k7kmw15af_WQsFmGNPDMkmkJyfqd9lELyNs5c&zhida_source=entity)来获得。\n 3. 通过全连接条件随机场来进行后处理，以改善分割结果。\n\n\n![](https://pica.zhimg.com/v2-e4e863acf8aff610b89598f3e8f7d28a_1440w.jpg)\n \n\n**Benchmarks (VOC2012)**\n![](https://pic2.zhimg.com/v2-737a292806d1557fa395ba06e17fe51b_1440w.jpg)\n## 5. RefineNet\n**论文信息**\n> RefineNet: Multi-Path Refinement Networks for High-Resolution Semantic Segmentation \n> Submitted on 20 Nov 2016 \n> [Arxiv Link](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/1611.06612)\n**主要贡献**\n 1. 精心设计了encoder-decoder架构中的decoder部分，使得性能提升。\n 2. 整个网络的设计都遵循[residual](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=residual&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJyZXNpZHVhbCIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjczMTE2ODgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.qz3TNrii00Juk8hNMprDsoM5FrdEuuLqq7mheUzpAsc&zhida_source=entity) connections，网络表达能力更强，[梯度](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E6%A2%AF%E5%BA%A6&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLmoq_luqYiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.TbF6pT3h6eTMYPwvzMxzVCpbpyPgquulkT_xr1CudJk&zhida_source=entity)更容易[反向传播](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E5%8F%8D%E5%90%91%E4%BC%A0%E6%92%AD&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLlj43lkJHkvKDmkq0iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.7MzFQMWhsdtWnmIru1OFLU9KmY4EX97BQW1lK8npzws&zhida_source=entity)。\n\n\n**简要概述**\n作者提出空洞卷积方法应用于语义分割也是有缺点的，包括：\n * 因为使用了大分辨率的[feature map](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=feature+map&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJmZWF0dXJlIG1hcCIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjczMTE2ODgsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.gRP8AFwXtjL-zUjp1dgUKdc4hlYt4efUpqPr85Kjvv0&zhida_source=entity)，因此计算代价大，并且需要大量的内存。对于这个问题，DeepLab的做法是只预测原始输入的1／8。\n\n\n本文提出使用encoder-decoder架构。encoder部分是RESNET-101。decoder具有RefineNet blocks，它将此前的RefineNet blocks的低分辨率特征和encoder部分[高分辨率](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E9%AB%98%E5%88%86%E8%BE%A8%E7%8E%87&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLpq5jliIbovqjnjociLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.I1XSsl6T9y0JdzgaVmrUkDhBmMjTcLI7uJ4ef9S4I_U&zhida_source=entity)特征进行concatenate/fuse。\n![](https://pic1.zhimg.com/v2-7c360b0719731484cc75ff11f2bc06e4_1440w.jpg)\n![](https://pic1.zhimg.com/v2-251ea8636e35e92cd886083bbce8f71e_1440w.jpg)\n**Benchmarks (VOC2012)**\n![](https://pic2.zhimg.com/v2-47a5f6788e7387cc76f6267d22c3b0cb_1440w.jpg)\n## 6. PSPNet\n**论文信息**\n> [Pyramid](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=Pyramid&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiJQeXJhbWlkIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6NzMxMTY4OCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.8cHFzuIcN9MZdbcJW3dzPG3-Y0VDEl3K_g7tgDg0nN0&zhida_source=entity) Scene Parsing Network \n> Submitted on 4 Dec 2016 \n> [Arxiv Link](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/1612.01105)\n**主要贡献**\n 1. 使用pyramid pooling整合context。\n 2. 使用auxiliary loss。\n\n\n**概要**\n[骨架网络](https://zhida.zhihu.com/search?content_id=7311688&content_type=Article&match_order=1&q=%E9%AA%A8%E6%9E%B6%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDcsInEiOiLpqqjmnrbnvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3MzExNjg4LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.Q-GB5AwmKw_jwzfIkDXBH6YjLmoom_Sq5wnlSepbV1M&zhida_source=entity)使用Resnet，并在此基础上加上pyramid pooling module。该模块用到了很多kernel大小不一的pooling 。将pooling的结果再上采样，经过concatenate进行融合。\n![](https://pic4.zhimg.com/v2-7c8c403f6c4f6962bda4780d6d27656f_1440w.jpg)\n在RESNET的第四阶段（即输入到金字塔池模块）之后，应用auxiliary loss。这种方法在别的地方也被称为intermediate supervision。\n**Benchmarks (VOC2012)**\n![](https://pic3.zhimg.com/v2-dc7a852a54823cecff329415a11ecabc_1440w.jpg)\n## 以下两篇论文还没细看，看完了之后更新。\n## 7. Large Kernel Matters\npass\n## 8. deeplab v3\npass\n \n\n**参考**\n这篇文章主要是参考以下两篇文章的一个总结，加入了一些自己的理解，想看原文的直接打开：\n 1. [A 2017 Guide to Semantic Segmentation with Deep Learning](https://link.zhihu.com/?target=http%3A//blog.qure.ai/notes/semantic-segmentation-deep-learning-review%23dilation)\n 2. [A Review on Deep Learning Techniques Applied to Semantic Segmentation](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/1704.06857)\n\n\n编辑于 2018-06-10 00:30\n[深度学习（Deep Learning）](https://www.zhihu.com/topic/19813032)\n[机器学习](https://www.zhihu.com/topic/19559450)\n[计算机视觉](https://www.zhihu.com/topic/19590195)\n​赞同 666​​19 条评论​1183 ​74 \n​分享\n​申请转载​\n​\n![](https://pic1.zhimg.com/v2-d41c2ceaed8f51999522f903672a521f_l.jpeg)\n未登录用户\n \n\n19 条评论\n默认\n最新\n[![dadream](https://pica.zhimg.com/v2-ad0d8ff83ee0aea2649c6c5efeb96256_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/008813f5db257ab17c34c20f970a2a43)\n[dadream](https://www.zhihu.com/people/008813f5db257ab17c34c20f970a2a43)\n没有unet？\n2018-09-20\n​回复​3\n展开其他 1 条回复​\n[![tan知yi](https://pic1.zhimg.com/v2-4accb568e8ce931d3ce317158d7acdf3_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/93e7e9bfdea1eafb71b842e2e45cd275)\n[tan知yi](https://www.zhihu.com/people/93e7e9bfdea1eafb71b842e2e45cd275)\n想知道传统的随机森林是如何做语义分割的![\\[谢邀\\]](https://pic2.zhimg.com/v2-6fe2283baa639ae1d7c024487f1d68c7.png)![\\[捂脸\\]](https://pic1.zhimg.com/v2-b62e608e405aeb33cd52830218f561ea.png)\n2020-12-28\n​回复​1\n[![寒青](https://picx.zhimg.com/v2-abed1a8c04700ba7d72b45195223e0ff_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/b2863314f4f336526ae871e34f21dfc7)\n[寒青](https://www.zhihu.com/people/b2863314f4f336526ae871e34f21dfc7)\n想问下，分割图不同类别不同颜色是如何实现的\n2020-07-04\n​回复​1\n展开其他 1 条回复​\n[![wong或许对了](https://picx.zhimg.com/v2-605c0d86ad0372497cca2b5dc60a1d47_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/956c37aa6e474f888ba4b689b8cd9e43)\n[wong或许对了](https://www.zhihu.com/people/956c37aa6e474f888ba4b689b8cd9e43)\n大佬不讲讲OCRnet吗\n2023-10-13\n​回复​喜欢\n[![大山](https://picx.zhimg.com/v2-76dfad8a3fff2fdd82d22d0b76e6d65b_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/a4304b214b7d2909d7dd0b69aa90a791)\n[大山](https://www.zhihu.com/people/a4304b214b7d2909d7dd0b69aa90a791)\n很棒的综述，非常谢谢！\n2022-10-18\n​回复​喜欢\n[![知乎用户dwCPx8](https://pic1.zhimg.com/v2-abed1a8c04700ba7d72b45195223e0ff_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/7bf35d9cf5bb04ea23b22572ea6c40a1)\n[知乎用户dwCPx8](https://www.zhihu.com/people/7bf35d9cf5bb04ea23b22572ea6c40a1)\n1\n2022-09-22\n​回复​喜欢\n[![Eric](https://picx.zhimg.com/v2-8f2c76c4c7c4934180f1aea67f20ca4f_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/3d0cac71d63596fd1d20c0223954020c)\n[Eric](https://www.zhihu.com/people/3d0cac71d63596fd1d20c0223954020c)\n![\\[赞同\\]](https://pic2.zhimg.com/v2-419a1a3ed02b7cfadc20af558aabc897.png)\n2021-06-02\n​回复​喜欢\n[![月色血风暴](https://picx.zhimg.com/v2-96c3a2246d9550547ebc9d174f40b0de_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/a18eaeb489cced34ef81da46c4cb694f)\n[月色血风暴](https://www.zhihu.com/people/a18eaeb489cced34ef81da46c4cb694f)\n感谢~写得很棒\n2020-02-21\n​回复​喜欢\n[![storm](https://picx.zhimg.com/v2-53e57f168827eb8f163b4c1eec764533_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/81cfaab7a1b8b5524328bc14f5feaf3b)\n[storm](https://www.zhihu.com/people/81cfaab7a1b8b5524328bc14f5feaf3b)\n感谢！可以更新吗\n2019-11-02\n​回复​喜欢\n[![李文鑫](https://pic1.zhimg.com/v2-ef96c661539a6586d3e2fbcb3f401e37_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/1d9439567b386e7df3e19d60d1bc1c33)\n[李文鑫](https://www.zhihu.com/people/1d9439567b386e7df3e19d60d1bc1c33)\n但是上次看论文 没看懂的 看了这篇之后 会更懂 谢谢作者了\n2019-09-25\n​回复​喜欢\n点击查看全部评论\n![](https://pic1.zhimg.com/v2-d41c2ceaed8f51999522f903672a521f_l.jpeg)\n未登录用户\n \n\n关于作者\n[![stone](https://picx.zhimg.com/36f310418_l.jpg?source=32738c0c&needBackground=1)](https://www.zhihu.com/people/big-big-stone)\n[stone](https://www.zhihu.com/people/big-big-stone)\n收藏之前点个赞呗\n[ 回答 **30** ](https://www.zhihu.com/people/big-big-stone/answers)[ 文章 **15** ](https://www.zhihu.com/people/big-big-stone/posts)[ 关注者 **10,013** ](https://www.zhihu.com/people/big-big-stone/followers)\n​关注他​发私信\n大家都在搜\n换一换\n[江歌妈妈发长文尘埃终将落定704 万](https://www.zhihu.com/search?q=%E6%B1%9F%E6%AD%8C%E5%A6%88%E5%A6%88%E5%8F%91%E9%95%BF%E6%96%87%E5%B0%98%E5%9F%83%E7%BB%88%E5%B0%86%E8%90%BD%E5%AE%9A&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)热\n[中国男足 0-5 巴勒斯坦562 万](https://www.zhihu.com/search?q=%E4%B8%AD%E5%9B%BD%E7%94%B7%E8%B6%B3+0-5+%E5%B7%B4%E5%8B%92%E6%96%AF%E5%9D%A6&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)热\n[原央视主持人阿丘回应被通报419 万](https://www.zhihu.com/search?q=%E5%8E%9F%E5%A4%AE%E8%A7%86%E4%B8%BB%E6%8C%81%E4%BA%BA%E9%98%BF%E4%B8%98%E5%9B%9E%E5%BA%94%E8%A2%AB%E9%80%9A%E6%8A%A5&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)热\n[华为与赛力斯达成新五年合作409 万](https://www.zhihu.com/search?q=%E5%8D%8E%E4%B8%BA%E4%B8%8E%E8%B5%9B%E5%8A%9B%E6%96%AF%E8%BE%BE%E6%88%90%E6%96%B0%E4%BA%94%E5%B9%B4%E5%90%88%E4%BD%9C&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)热\n[迪拜航空确认航班发生事故398 万](https://www.zhihu.com/search?q=%E8%BF%AA%E6%8B%9C%E8%88%AA%E7%A9%BA%E7%A1%AE%E8%AE%A4%E8%88%AA%E7%8F%AD%E5%8F%91%E7%94%9F%E4%BA%8B%E6%95%85&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)热\n[东航回应网传空姐跪地道歉326 万](https://www.zhihu.com/search?q=%E4%B8%9C%E8%88%AA%E5%9B%9E%E5%BA%94%E7%BD%91%E4%BC%A0%E7%A9%BA%E5%A7%90%E8%B7%AA%E5%9C%B0%E9%81%93%E6%AD%89&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)\n[王楚钦：不知道为什么就是感觉累320 万](https://www.zhihu.com/search?q=%E7%8E%8B%E6%A5%9A%E9%92%A6%EF%BC%9A%E4%B8%8D%E7%9F%A5%E9%81%93%E4%B8%BA%E4%BB%80%E4%B9%88%E5%B0%B1%E6%98%AF%E6%84%9F%E8%A7%89%E7%B4%AF&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)热\n[江苏高考作文《衬衫的价格为 9 磅 15 便士》爆火268 万](https://www.zhihu.com/search?q=%E6%B1%9F%E8%8B%8F%E9%AB%98%E8%80%83%E4%BD%9C%E6%96%87%E3%80%8A%E8%A1%AC%E8%A1%AB%E7%9A%84%E4%BB%B7%E6%A0%BC%E4%B8%BA+9+%E7%A3%85+15+%E4%BE%BF%E5%A3%AB%E3%80%8B%E7%88%86%E7%81%AB&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)\n[女装网店开始用防拆带了265 万](https://www.zhihu.com/search?q=%E5%A5%B3%E8%A3%85%E7%BD%91%E5%BA%97%E5%BC%80%E5%A7%8B%E7%94%A8%E9%98%B2%E6%8B%86%E5%B8%A6%E4%BA%86&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)\n[孕妇骑车别车被司机踹翻257 万](https://www.zhihu.com/search?q=%E5%AD%95%E5%A6%87%E9%AA%91%E8%BD%A6%E5%88%AB%E8%BD%A6%E8%A2%AB%E5%8F%B8%E6%9C%BA%E8%B8%B9%E7%BF%BB&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=7e924ff0-c2b6-4a5c-998a-656e894c6012)\n### 推荐阅读\n# [浅谈图像的语义分割 基于深度学习的图像语义分割综述 摘要 近些年来计算机视觉与机器学习研究者对图像语义分割问题越来越感兴趣，越来越多的应用场景需要精确且高效的分割技术，如自动驾驶、室内导航、甚至虚拟… Laayt ](https://zhuanlan.zhihu.com/p/100140440)[ ![图片语义分割深度学习算法要点回顾](https://picx.zhimg.com/v2-f70c8096d33ac457b2fbd4c9b4af1c7d_250x0.jpg?source=172ae18b)图片语义分割深度学习算法要点回顾 AI研习社 ](https://zhuanlan.zhihu.com/p/53233072)[ ![图像语义分割\\(20\\) 通过图像合成方法检测训练中未出现的类别未知的物体](https://picx.zhimg.com/v2-ce9e966f9a400b569d49e6d1bfacd616_250x0.jpg?source=172ae18b)图像语义分割(20) 通过图像合成方法检测训练中未出现的类别未知的物体 赵磊发表于计算机科学... ](https://zhuanlan.zhihu.com/p/78193038)[ ![【图解AI】什么是语义分割、实例分割、全景分割](https://picx.zhimg.com/v2-ef8c1f79c6a03c5bfe270cd8bd9d3678_250x0.jpg?source=172ae18b)【图解AI】什么是语义分割、实例分割、全景分割 Jack Wang ](https://zhuanlan.zhihu.com/p/368904941) \n\n_想来知乎工作？请发送邮件到 jobs@zhihu.com_\n打开知乎App\n在「我的页」右上角打开扫一扫\n![](https://picx.zhimg.com/v2-9e41ea16bdfbe9cf4896617ecad5b4ca.png)\n其他扫码方式：微信\n下载知乎App\n无障碍模式\n验证码登录\n密码登录\n[开通机构号](https://zhuanlan.zhihu.com/org/signup)\n中国 +86\n获取短信验证码\n忘记密码\n登录/注册\n其他方式登录\n未注册手机验证后自动登录，注册即代表同意[《知乎协议》](https://www.zhihu.com/term/zhihu-terms)[《隐私保护指引》](https://www.zhihu.com/term/privacy)\n扫码下载知乎 App\n关闭二维码\n", "body_base_path": "attachments/dd6eafdfdf38/pages/", "body_markdown_path": "attachments/dd6eafdfdf38/pages/archived-body.md", "local_html": "attachments/dd6eafdfdf38/pages/archived-body.html", "body_source": "prior_capture" }