{"id":6522,"date":"2022-08-30T20:42:00","date_gmt":"2022-08-30T12:42:00","guid":{"rendered":"http:\/\/139.9.1.231\/?p=6522"},"modified":"2022-08-28T20:44:15","modified_gmt":"2022-08-28T12:44:15","slug":"segnet","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/08\/30\/segnet\/","title":{"rendered":"SegNet"},"content":{"rendered":"\n<p>\u8bba\u6587\uff082015\uff09\uff1a<a rel=\"noreferrer noopener\" href=\"https:\/\/arxiv.org\/abs\/1511.00561\" target=\"_blank\">SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation<\/a><\/p>\n\n\n\n<p> Github\uff1a<a href=\"https:\/\/github.com\/alexgkendall\/caffe-segnet\">https:\/\/github.com\/alexgkendall\/caffe-segnet<\/a><\/p>\n\n\n\n<p>                    <\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full\"><img loading=\"lazy\" width=\"537\" height=\"149\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-410.png\" alt=\"\" class=\"wp-image-6525\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-410.png 537w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-410-300x83.png 300w\" sizes=\"(max-width: 537px) 100vw, 537px\" \/><\/figure><\/div>\n\n\n\n<ul><li>\u628a\u672c\u6587\u63d0\u51fa\u7684\u67b6\u6784\u548cFCN\u3001DeepLab-LargeFOV\u3001DeconvNet\u505a\u4e86\u6bd4\u8f83\uff0c\u8fd9\u79cd\u6bd4\u8f83\u63ed\u793a\u4e86\u5728\u5b9e\u73b0\u826f\u597d\u5206\u5272\u6027\u80fd\u7684\u524d\u63d0\u4e0b\u5185\u5b58\u4f7f\u7528\u60c5\u51b5\u4e0e\u5206\u5272\u51c6\u786e\u6027\u7684\u6743\u8861\u3002<\/li><li>SegNet\u7684\u4e3b\u8981\u52a8\u673a\u662f\u573a\u666f\u7406\u89e3\u7684\u5e94\u7528\u3002\u56e0\u6b64\u5b83\u5728\u8bbe\u8ba1\u7684\u65f6\u5019\u8003\u8651\u4e86\u8981\u5728\u9884\u6d4b\u671f\u95f4\u4fdd\u8bc1\u5185\u5b58\u548c\u8ba1\u7b97\u65f6\u95f4\u4e0a\u7684\u6548\u7387\u3002<\/li><li>\u5b9a\u91cf\u7684\u8bc4\u4f30\u8868\u660e\uff0cSegNet\u5728\u548c\u5176\u4ed6\u67b6\u6784\u7684\u6bd4\u8f83\u4e0a\uff0c\u65f6\u95f4\u548c\u5185\u5b58\u7684\u4f7f\u7528\u90fd\u6bd4\u8f83\u9ad8\u6548\u3002<\/li><\/ul>\n\n\n\n<p>      SegNet\u8bba\u6587\u63d0\u51fa\u4e86max pooling\u7684\u6539\u8fdb\u7248\uff0c\u4f7f\u7528\u8be5pooling\u64cd\u4f5c\u65e2\u53ef\u4ee5\u8fdb\u884c\u4e0b\u91c7\u6837\u64cd\u4f5c\uff0c\u4e5f\u53ef\u4ee5\u8fdb\u884c\u4e0a\u91c7\u6837\u64cd\u4f5c\u3002\u5728\u4e0b\u91c7\u6837\u64cd\u4f5c\u4e2d\u540c\u65f6\u8f93\u51fapooling\u540e\u7684\u7ed3\u679c\u548cpooling\u8fc7\u7a0b\u4e2d\u7684\u7d22\u5f15\u3002\u5728\u4e0a\u91c7\u6837\u64cd\u4f5c\u4e2d\uff0c\u5229\u7528\u4e0b\u91c7\u6837\u5bf9\u5e94\u4f4d\u7f6e\u7684\u7d22\u5f15\uff0c\u8fdb\u884c\u4e0a\u91c7\u6837\u64cd\u4f5c\uff0c\u8fd9\u6837\u7684\u4f18\u52bf\u5728\u4e8e\u8bb0\u4f4f\u4e86\u6700\u4eae\u7279\u5f81\u50cf\u7d20\u7684\u7a7a\u95f4\u4f4d\u7f6e\u3002\uff08\u53bb\u9664\u4e86unet\u91cc\u9762\u7684\u53cd\u5377\u79ef\u64cd\u4f5c\uff09<\/p>\n\n\n\n<p>\u4f18\u70b9\uff0c<\/p>\n\n\n\n<ol><li>\u53ef\u4ee5\u63d0\u9ad8\u7269\u4f53\u8fb9\u754c\u7684\u5206\u5272\u6548\u679c<\/li><li>\u76f8\u6bd4\u53cd\u5377\u79ef\u64cd\u4f5c\uff0c\u51cf\u5c11\u4e86\u53c2\u6570\u6570\u91cf\uff0c\u51cf\u5c11\u4e86\u8fd0\u7b97\u91cf\uff0c\u76f8\u6bd4resize\u64cd\u4f5c\uff0c\u51cf\u5c11\u4e86\u63d2\u503c\u7684\u8fd0\u7b97\u91cf\uff0c\u800c\u5b9e\u9645\u589e\u52a0\u7684\u7d22\u5f15\u53c2\u6570\u4e5f\u5f88\u5c11\u3002<\/li><li>\u8be5pooling\u64cd\u4f5c\u53ef\u4ee5\u5e94\u7528\u4e8e\u4efb\u4f55\u57fa\u4e8e\u7f16\u7801-\u89e3\u7801\u7684\u5206\u5272\u6a21\u578b\u3002<\/li><\/ol>\n\n\n\n<p>        SegNet\u7f51\u7edc\u7ed3\u6784\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u662f\u4e00\u4e2a\u7f16\u89e3\u7801\u5b8c\u5168\u5bf9\u79f0\u7684\u7ed3\u6784\u3002\u5176\u7f16\u7801\u5668\u76f4\u63a5\u7528\u4e86VGG16\u7684\u7ed3\u6784\uff0c\u5e76\u5c06\u5168\u8fde\u63a5\u5c42\u5168\u90e8\u6539\u4e3a\u5377\u79ef\u5c42\uff0c\u5b9e\u9645\u8bad\u7ec3\u65f6\u53ef\u4f7f\u7528VGG16\u7684\u9884\u8bad\u7ec3\u6743\u91cd\u8fdb\u884c\u521d\u59cb\u5316\uff1b\u7f16\u7801\u5668\u5c0613\u5c42\u5377\u79ef\u5c42\u5206\u4e3a5\u7ec4\u5377\u79ef\u5757\uff0c\u6bcf\u7ec4\u5377\u79ef\u5757\u4e4b\u95f4\u7528\u6700\u5927\u6c60\u5316\u5c42\u8fdb\u884c\u4e0b\u91c7\u6837\u3002\u4f5c\u4e3a\u4e00\u4e2a\u5bf9\u79f0\u7ed3\u6784\uff0cSegNet\u89e3\u7801\u5668\u4e5f\u670913\u5c42\u5377\u79ef\u5c42\uff0c\u540c\u6837\u5206\u4e3a5\u7ec4\u5377\u79ef\u5757\uff0c\u6bcf\u7ec4\u5377\u79ef\u5757\u4e4b\u95f4\u7528\u53cc\u7ebf\u6027\u63d2\u503c\u548c\u6700\u5927\u6c60\u5316\u4f4d\u7f6e\u7d22\u5f15\u8fdb\u884c\u4e0a\u91c7\u6837\uff0c\u8fd9\u4e5f\u662fSegNet\u6700\u5927\u7684\u7279\u8272\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"763\" height=\"288\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-411.png\" alt=\"\" class=\"wp-image-6530\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-411.png 763w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-411-300x113.png 300w\" sizes=\"(max-width: 763px) 100vw, 763px\" \/><\/figure>\n\n\n\n<p>       SegNet\u7814\u7a76\u56e2\u961f\u8ba4\u4e3a\u7f16\u7801\u5668\u4e0b\u91c7\u6837\u8fc7\u7a0b\u4e2d\u56fe\u50cf\u4fe1\u606f\u635f\u5931\u8f83\u591a\uff0c\u76f4\u63a5\u5b58\u50a8\u6240\u6709\u5377\u79ef\u5757\u7684\u7279\u5f81\u56fe\u53c8\u975e\u5e38\u5360\u7528\u5185\u5b58\uff0c\u56e0\u800c\u5728SegNet\u4e2d\u63d0\u51fa\u5728\u6bcf\u4e00\u6b21\u6700\u5927\u6c60\u5316\u4e0b\u91c7\u6837\u524d\u5b58\u50a8\u6700\u5927\u6c60\u5316\u7684\u4f4d\u7f6e\u7d22\u5f15\uff08Max-pooling indices\uff09\uff0c\u5373\u8bb0\u4f4f\u6700\u5927\u6c60\u5316\u64cd\u4f5c\u4e2d\uff0c\u6700\u5927\u503c\u57282*2\u6c60\u5316\u7a97\u53e3\u4e2d\u7684\u4f4d\u7f6e\u3002\u6bcf\u4e2a2*2\u7a97\u53e3\u4ec5\u9700\u89812 bits\u5185\u5b58\u5b58\u50a8\u91cf\uff0c\u8fd9\u79cd\u6c60\u5316\u4f4d\u7f6e\u7d22\u5f15\u53ef\u7528\u4e8e\u4e0a\u91c7\u6837\u89e3\u7801\u65f6\u6062\u590d\u56fe\u50cf\u4fe1\u606f\u3002\u4e0b\u56fe\u7ed9\u51fa\u4e86SegNet\u4e0eFCN\u4e4b\u95f4\u7684\u4e0a\u91c7\u6837\u65b9\u6cd5\u5bf9\u6bd4\u3002\u53ef\u4ee5\u89c2\u5bdf\u5230\uff0cSegNet\u4f7f\u7528\u53cc\u7ebf\u6027\u63d2\u503c\u5e76\u7ed3\u5408\u6700\u5927\u6c60\u5316\u4f4d\u7f6e\u7d22\u5f15\u8fdb\u884c\u4e0a\u91c7\u6837\uff0c\u800cFCN\u5219\u662f\u57fa\u4e8e\u53bb\u5377\u79ef\u7ed3\u5408\u7f16\u7801\u5668\u5377\u79ef\u7279\u5f81\u56fe\u8fdb\u884c\u4e0a\u91c7\u6837\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"711\" height=\"320\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-412.png\" alt=\"\" class=\"wp-image-6531\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-412.png 711w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-412-300x135.png 300w\" sizes=\"(max-width: 711px) 100vw, 711px\" \/><\/figure>\n\n\n\n<p>     SegNet\u8fd9\u79cd\u8f7b\u91cf\u5316\u7684\u4e0a\u91c7\u6837\u65b9\u5f0f\uff0c\u4e0d\u4ec5\u80fd\u591f\u63d0\u5347\u56fe\u50cf\u8fb9\u754c\u5206\u5272\u6548\u679c\uff0c\u5728\u7aef\u5230\u7aef\u7684\u5b9e\u65f6\u5206\u5272\u9879\u76ee\u4e2d\u901f\u5ea6\u4e5f\u975e\u5e38\u5feb\uff0c\u5e76\u4e14\u8fd9\u79cd\u7ed3\u6784\u8bbe\u8ba1\u53ef\u4ee5\u914d\u7f6e\u5230\u4efb\u610f\u7684\u7f16\u89e3\u7801\u7f51\u7edc\u4e2d\uff0c\u662f\u4e00\u79cd\u4f18\u79c0\u7684\u5206\u5272\u7f51\u7edc\u8bbe\u8ba1\u65b9\u5f0f\u3002\u4e0b\u8ff0\u4ee3\u7801\u7ed9\u51fa\u4e86SegNet\u7684\u4e00\u4e2a\u7b80\u6613\u7684\u7ed3\u6784\u5b9e\u73b0\uff0c\u56e0\u4e3aSegNet\u89e3\u7801\u5668\u7684\u7279\u6b8a\u6027\uff0c\u6211\u4eec\u5355\u72ec\u5b9a\u4e49\u4e86\u4e00\u4e2a\u89e3\u7801\u5668\u7c7b\uff0c\u7f16\u7801\u5668\u90e8\u5206\u76f4\u63a5\u4f7f\u7528VGG16\u7684\u9884\u8bad\u7ec3\u6743\u91cd\u5c42\uff0c\u7136\u540e\u5728\u7f16\u89e3\u7801\u5668\u57fa\u7840\u4e0a\u642d\u5efaSegNet\u5e76\u5b9a\u4e49\u524d\u5411\u8ba1\u7b97\u6d41\u7a0b\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>\r\n# \u5bfc\u5165PyTorch\u76f8\u5173\u6a21\u5757\r\nimport torch\r\nimport torch.nn as nn\r\nimport torch.nn.init as init\r\nimport torch.nn.functional as F\r\nfrom torchvision import models\r\n\r\n# \u5b9a\u4e49SegNet\u89e3\u7801\u5668\u7c7b\r\nclass SegNetDec(nn.Module):\r\n    def __init__(self, in_channels, out_channels, num_layers):\r\n        super().__init__()\r\n        layers = &#91;\r\n            nn.Conv2d(in_channels, in_channels \/\/ 2, 3, padding=1),\r\n            nn.BatchNorm2d(in_channels \/\/ 2),\r\n            nn.ReLU(inplace=True),\r\n        ]\r\n        layers += &#91;\r\n            nn.Conv2d(in_channels \/\/ 2, in_channels \/\/ 2, 3, padding=1),\r\n            nn.BatchNorm2d(in_channels \/\/ 2),\r\n            nn.ReLU(inplace=True),\r\n        ] * num_layers\r\n        layers += &#91;\r\n            nn.Conv2d(in_channels \/\/ 2, out_channels, 3, padding=1),\r\n            nn.BatchNorm2d(out_channels),\r\n            nn.ReLU(inplace=True),\r\n        ]\r\n        self.decode = nn.Sequential(*layers)\r\n\r\n    def forward(self, x):\r\n        return self.decode(x)\r\n\r\n### \u5b9a\u4e49SegNet\u7c7b\r\nclass SegNet(nn.Module):\r\n    def __init__(self, classes):\r\n        super().__init__()\r\n    # \u7f16\u7801\u5668\u4f7f\u7528vgg16\u9884\u8bad\u7ec3\u6743\u91cd\r\n        vgg16 = models.vgg16(pretrained=True)\r\n        features = vgg16.features\r\n        self.enc1 = features&#91;0: 4]\r\n        self.enc2 = features&#91;5: 9]\r\n        self.enc3 = features&#91;10: 16]\r\n        self.enc4 = features&#91;17: 23]\r\n        self.enc5 = features&#91;24: -1]\r\n    # \u7f16\u7801\u5668\u5377\u79ef\u5c42\u4e0d\u53c2\u4e0e\u8bad\u7ec3\r\n        for m in self.modules():\r\n            if isinstance(m, nn.Conv2d):\r\n                m.requires_grad = False\r\n    \r\n        self.dec5 = SegNetDec(512, 512, 1)\r\n        self.dec4 = SegNetDec(512, 256, 1)\r\n        self.dec3 = SegNetDec(256, 128, 1)\r\n        self.dec2 = SegNetDec(128, 64, 0)\r\n\r\n        self.final = nn.Sequential(*&#91;\r\n            nn.Conv2d(64, classes, 3, padding=1),\r\n            nn.BatchNorm2d(classes),\r\n            nn.ReLU(inplace=True)\r\n        ])\r\n  # \u5b9a\u4e49SegNet\u524d\u5411\u8ba1\u7b97\u6d41\u7a0b\r\n    def forward(self, x):\r\n        x1 = self.enc1(x)\r\n        e1, m1 = F.max_pool2d(x1, kernel_size=2, stride=2,\r\n return_indices=True)\r\n        x2 = self.enc2(e1)\r\n        e2, m2 = F.max_pool2d(x2, kernel_size=2, stride=2,\r\n return_indices=True)\r\n        x3 = self.enc3(e2)\r\n        e3, m3 = F.max_pool2d(x3, kernel_size=2, stride=2,\r\n return_indices=True)\r\n        x4 = self.enc4(e3)\r\n        e4, m4 = F.max_pool2d(x4, kernel_size=2, stride=2,\r\n return_indices=True)\r\n        x5 = self.enc5(e4)\r\n        e5, m5 = F.max_pool2d(x5, kernel_size=2, stride=2,\r\n return_indices=True)\r\n\r\n        def upsample(d):\r\n            d5 = self.dec5(F.max_unpool2d(d, m5, kernel_size=2,\r\n stride=2, output_size=x5.size()))\r\n            d4 = self.dec4(F.max_unpool2d(d5, m4, kernel_size=2,\r\n stride=2, output_size=x4.size()))\r\n            d3 = self.dec3(F.max_unpool2d(d4, m3, kernel_size=2,\r\n stride=2, output_size=x3.size()))\r\n            d2 = self.dec2(F.max_unpool2d(d3, m2, kernel_size=2,\r\n stride=2, output_size=x2.size()))\r\n            d1 = F.max_unpool2d(d2, m1, kernel_size=2, stride=2,\r\n output_size=x1.size())\r\n            return d1\r\n\r\n        d = upsample(e5)\r\n        return self.final(d)<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587\uff082015\uff09\uff1aSegNet: A Deep Convolutional Encoder-Decoder A &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2022\/08\/30\/segnet\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">SegNet<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[24],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/6522"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=6522"}],"version-history":[{"count":10,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/6522\/revisions"}],"predecessor-version":[{"id":6535,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/6522\/revisions\/6535"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=6522"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=6522"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=6522"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}