{"id":6536,"date":"2022-08-31T22:11:52","date_gmt":"2022-08-31T14:11:52","guid":{"rendered":"http:\/\/139.9.1.231\/?p=6536"},"modified":"2022-08-31T22:11:52","modified_gmt":"2022-08-31T14:11:52","slug":"refinenet","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/08\/31\/refinenet\/","title":{"rendered":"RefineNet"},"content":{"rendered":"\n

\u8bba\u6587\u5730\u5740\uff082016\uff09\uff1aRefineNet: Multi-Path Refinement Networks with Identity Mappings for High-Resolution Semantic Segmentation<\/a><\/p>\n\n\n\n

\"\"<\/figure><\/div>\n\n\n\n

\u5bf9\u4e8e\u9ad8\u5206\u8fa8\u7387\u7684\u56fe\u50cf\u5206\u5272\u95ee\u9898\uff0c\u57fa\u4e8e\u7f16\u89e3\u7801\u7ed3\u6784\u7684\u5206\u5272\u7f51\u7edc\u867d\u7136\u6709\u6548\uff0c\u4f46\u56e0\u4e3a\u5377\u79ef\u548c\u6c60\u5316\u4e0b\u91c7\u6837\u7684\u5b58\u5728\uff0c\u7279\u5f81\u56fe\u5728\u53d8\u5c0f\u7684\u8fc7\u7a0b\u4f1a\u9010\u6e10\u635f\u5931\u4e00\u4e9b\u7ec6\u7c92\u5ea6\u7684\u4fe1\u606f\uff0c\u975e\u5e38\u4e0d\u5229\u4e8e\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u7684\u50cf\u7d20\u7a20\u5bc6\u9884\u6d4b\u3002\u9488\u5bf9\u8fd9\u4e2a\u95ee\u9898\uff0c\u6b64\u524d\u7684\u5404\u9879\u7814\u7a76\u5f52\u7eb3\u800c\u8a00\u63d0\u51fa\u4e86\u5982\u4e0b\u4e09\u70b9\u5904\u7406\u65b9\u6cd5\uff1a<\/p>\n\n\n\n

\uff081\uff09\u7c7b\u4f3c\u4e8eFCN\u548cUNet\uff0c\u76f4\u63a5\u4f7f\u7528\u8f6c\u7f6e\u5377\u79ef\u4e0a\u91c7\u6837\u6765\u6062\u590d\u56fe\u50cf\u50cf\u7d20\uff0c\u4f46\u8f6c\u7f6e\u5377\u79ef\u5bf9\u4e8e\u4e0b\u91c7\u6837\u8fc7\u7a0b\u4e2d\u4e22\u5931\u7684\u4f4e\u5c42\u4fe1\u606f\u7684\u6062\u590d\u80fd\u529b\u6709\u9650\u3002<\/p>\n\n\n\n

\uff082\uff09\u4f7f\u7528\u7a7a\u6d1e\u5377\u79ef\uff0c\u901a\u8fc7\u7ed9\u5e38\u89c4\u5377\u79ef\u4e2d\u63d2\u5165\u7a7a\u6d1e\u7684\u65b9\u5f0f\u6765\u589e\u5927\u5377\u79ef\u611f\u53d7\u91ce\uff0c\u5e76\u4e14\u6ca1\u6709\u7f29\u5c0f\u56fe\u50cf\u5c3a\u5bf8\uff0c\u4f46\u8fd9\u79cd\u65b9\u5f0f\u8ba1\u7b97\u5f00\u9500\u589e\u5927\uff0c\u6a21\u578b\u8fd0\u884c\u6548\u7387\u964d\u4f4e\uff0c\u5e76\u4e14\u7a7a\u6d1e\u5377\u79ef\u4f5c\u4e3a\u4e00\u79cd\u8f83\u4e3a\u7c97\u7cd9\u7684\u5b50\u91c7\u6837\uff08sub-sampling\uff09\uff0c\u4e5f\u4f1a\u5b58\u5728\u56fe\u50cf\u91cd\u8981\u4fe1\u606f\u635f\u5931\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n

\uff083\uff09\u4f7f\u7528\u8df3\u8dc3\u8fde\u63a5\u3002\u7c7b\u4f3c\u4e8eUNet\u4e2d\u7f16\u89e3\u7801\u5668\u95f4\u7684\u8df3\u8dc3\u8fde\u63a5\uff0c\u76f4\u63a5\u5c06\u7f16\u7801\u5668\u6bcf\u4e00\u5c42\u7684\u7279\u5f81\u56fe\u8fde\u63a5\u5230\u89e3\u7801\u5668\u4e0a\u91c7\u6837\u7ed3\u679c\u4e0a\uff0c\u80fd\u591f\u5bf9\u89e3\u7801\u56fe\u50cf\u8fdb\u884c\u4fe1\u606f\u8865\u5145\u3002<\/p>\n\n\n\n

\u76f8\u5173\u7814\u7a76\u8ba4\u4e3a\uff0c\u5bf9\u4e8e\u7f16\u89e3\u7801\u7ed3\u6784\u800c\u8a00\uff0c\u6240\u6709\u5c42\u6b21\u7684\u7279\u5f81\u5bf9\u8bed\u4e49\u5206\u5272\u90fd\u662f\u6709\u5e2e\u52a9\u7684\u3002\u9ad8\u5c42\u6b21\u7684\u7279\u5f81\u7528\u4e8e\u8bc6\u522b\u56fe\u50cf\u4e2d\u7684\u8bed\u4e49\u4fe1\u606f\uff0c\u4f4e\u5c42\u6b21\u7684\u7279\u5f81\u5219\u6709\u52a9\u4e8e\u6062\u590d\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u7684\u8fb9\u754c\u7ec6\u8282\u3002\u4f46\u5982\u4f55\u6709\u6548\u5229\u7528\u4e2d\u95f4\u5c42\u6b21\u7684\u4fe1\u606f\u503c\u5f97\u8fdb\u4e00\u6b65\u63a2\u7d22\uff0c\u524d\u8ff0\u5145\u5206\u4f7f\u7528\u8df3\u8dc3\u8fde\u63a5\u7684\u65b9\u6cd5\u6216\u8bb8\u4f1a\u66f4\u52a0\u6709\u6548\u3002\u57fa\u4e8e\u6b64\uff0c\u7814\u7a76\u4eba\u5458\u63d0\u51fa\u4e86\u4e00\u79cd\u9488\u5bf9\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u8bed\u4e49\u5206\u5272\u7684\u591a\u5c42\u6b21\u7279\u5f81\u7cbe\u7ec6\u5316\u7f51\u7edc\uff1aRefineNet\u3002\u63d0\u51faRefineNet\u7684\u8bba\u6587\u4e3aRefineNet: Multi-Path Refinement Networks for High-Resolution Semantic Segmentation\uff0c\u8be5\u7f51\u7edc\u57fa\u4e8eResNet\u7ed3\u6784\u548c\u8df3\u8dc3\u8fde\u63a5\uff0c\u4f7f\u7528\u591a\u8def\u5f84\u7684\u7cbe\u7ec6\u5316\u7f51\u7edc\u7ed3\u6784\u6765\u83b7\u53d6\u6700\u4f73\u7684\u5206\u5272\u7ed3\u679c\uff0c\u662f\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u5206\u5272\u7684\u7ecf\u5178\u7f51\u7edc\u3002<\/p>\n\n\n\n

RefineNet\u7b80\u8981\u7ed3\u6784\u5982\u56fe5-9\u6240\u793a\u3002RefineNet\u603b\u4f53\u4e0a\u4ecd\u7136\u662f\u7f16\u89e3\u7801\u7ed3\u6784\uff0c\u7f16\u7801\u5668\u90e8\u5206\u6839\u636e\u9884\u8bad\u7ec3\u7684ResNet\u7f51\u7edc\u5212\u5206\u4e864\u4e2a\u5377\u79ef\u5757\uff0c\u4e0e\u4e4b\u5bf9\u5e94\u7684\u662f\u7ea7\u8054\u4e864\u4e2aRefineNet\u5355\u5143\u5230\u89e3\u7801\u5668\u90e8\u5206\u3002\u6bcf\u4e2a\u7f16\u7801\u5668ResNet\u5377\u79ef\u5757\u7684\u8f93\u51fa\u7279\u5f81\u56fe\u90fd\u4f1a\u88ab\u8fde\u63a5\u5230\u5bf9\u5e94\u7684RefineNet\u5355\u5143\uff0c\u5982\u56fe\u4e2d\u7684ResNet-4\u8fde\u63a5\u5230RefineNet-4\u5355\u5143\uff0c\u5230\u4e86RefineNet-3\u5355\u5143\uff0c\u9664\u4e86\u63a5\u6536\u6765\u81eaResNet-3\u7684\u8f93\u51fa\u5916\uff0c\u8fd8\u9700\u8981\u63a5\u6536RefineNet-4\u5355\u5143\u7684\u8f93\u51fa\uff0c\u5bf9\u4e8eRefineNet-3\u5355\u5143\u800c\u8a00\u5c31\u6784\u6210\u4e86\u4e24\u8def\u5f84\u7684\u8f93\u5165\u3002\u8fd9\u6837\u5c42\u5c42\u5411\u4e0a\u7ea7\u8054\uff0c\u5c31\u6784\u6210\u4e86\u591a\u8def\u5f84\u7684RefineNet\u3002\u5176\u4e2d\u7f16\u7801\u5668\u4e2d\u6bcf\u4e2a\u7279\u5f81\u56fe\u5230\u89e3\u7801\u5668RefineNet\u5355\u5143\u7684\u8fde\u63a5\u4e5f\u53eb\u957f\u7a0b\u6b8b\u5dee\u8fde\u63a5\uff08long-range residual connections\uff09\u3002<\/p>\n\n\n\n

\"\"<\/figure><\/div>\n\n\n\n

\u4e0a\u56fe\u4ec5\u7ed9\u51fa\u4e86\u5305\u542b\u4e86\u7f16\u7801\u5668\u5728\u5185\u7684RefineNet\u7b80\u8981\u7ed3\u6784\uff0c\u800cRefineNet\u5355\u5143\u7684\u5177\u4f53\u7ed3\u6784\u5982\u4e0b\u56fe\u6240\u793a\u3002\u4e00\u4e2aRefineNet\u5355\u5143\u7531\u6b8b\u5dee\u5377\u79ef\u5355\u5143\uff08Residual convolution unit\uff0cRCU\uff09\u3001\u591a\u5206\u8fa8\u7387\u878d\u5408\uff08Multi-resolution Fusion\uff09\u548c\u94fe\u5f0f\u6b8b\u5dee\u6c60\u5316\uff08Chained Residual Pooling\uff0cCRP\uff09\u7ec4\u6210\u3002RCU\u8f83\u4e3a\u7b80\u5355\uff0c\u5c31\u662f\u5e38\u89c4\u7684ResNet\u7ed3\u6784\uff0c\u6bcf\u4e00\u4e2a\u8f93\u5165\u8def\u5f84\u90fd\u4f1a\u7ecf\u8fc7\u4e24\u6b21RCU\u64cd\u4f5c\u540e\u518d\u8f93\u51fa\u5230\u4e0b\u4e00\u4e2a\u5355\u5143\u3002RCU\u7684\u8df3\u8dc3\u8fde\u63a5\u5728RefineNet\u4e2d\u4e5f\u88ab\u79f0\u4e3a\u77ed\u7a0b\u6b8b\u5dee\u8fde\u63a5\uff08short-range residual connections\uff09\u3002\u7d27\u63a5\u7740\u662f\u4e00\u4e2a\u591a\u5206\u8fa8\u7387\u7279\u5f81\u56fe\u878d\u5408\u5c42\uff0c\u5c06\u4e0a\u4e00\u5c42RCU\u8f93\u51fa\u7684\u591a\u8def\u5f84\u7279\u5f81\u56fe\u7ecf\u8fc7\u4e00\u4e2a33\u7684\u5377\u79ef\u548c\u4e0a\u91c7\u6837\u64cd\u4f5c\u540e\u8fdb\u884c\u52a0\u603b\uff0c\u5f97\u5230\u5408\u5e76\u540e\u7684\u7279\u5f81\u56fe\u3002\u6700\u540e\u662f\u4e00\u4e2aCRP\u5355\u5143\uff0c\u8fd9\u4e5f\u662fRefineNet\u7684\u7279\u8272\u7ed3\u6784\uff0c\u901a\u8fc73\u4e2a\u94fe\u5f0f\u7684\u6c60\u5316\u548c\u5377\u79ef\u6b8b\u5dee\u7ec4\u5408\u6765\u6355\u6349\u5927\u56fe\u50cf\u533a\u57df\u7684\u80cc\u666f\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002\u5c06CRP\u4e4b\u540e\u5f97\u5230\u7279\u5f81\u56fe\u518d\u7ecf\u8fc7\u4e00\u6b21RCU\u5373\u53ef\u5230\u6700\u7ec8\u7684\u5206\u5272\u8f93\u51fa\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u4f5c\u4e3a\u4e00\u79cd\u9488\u5bf9\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u7684\u7cbe\u7ec6\u5316\u5206\u5272\u7f51\u7edc\uff0cRefineNet\u7684\u7ed3\u6784\u8bbe\u8ba1\u65e0\u7591\u662f\u6210\u529f\u7684\uff0c\u5f53\u65f6\u5728\u591a\u4e2a\u516c\u5f00\u6570\u636e\u96c6\u4e0a\u5747\u53d6\u5f97\u4e86SOTA\u6027\u80fd\u8868\u73b0\u3002\u8fd9\u79cd\u591a\u8def\u5f84\u7684\u7cbe\u7ec6\u5316\u7f51\u7edc\u80fd\u591f\u901a\u8fc7\u8fed\u4ee3\u7cbe\u70bc\u7684\u65b9\u5f0f\u5c06\u7c97\u7cd9\u7684\u8bed\u4e49\u7279\u5f81\u7cbe\u70bc\u4e3a\u7ec6\u7c92\u5ea6\u7684\u8bed\u4e49\u7279\u5f81\u3002\u5176\u6b21\uff0c\u57fa\u4e8e\u957f\u77ed\u7a0b\u7684\u6b8b\u5dee\u8fde\u63a5\u80fd\u591f\u4f7f\u5f97\u6a21\u578b\u8fdb\u884c\u7aef\u5230\u7aef\u7684\u8bad\u7ec3\uff0c\u63a8\u7406\u65f6\u4e5f\u975e\u5e38\u9ad8\u6548\u3002\u6700\u540e\uff0c\u94fe\u5f0f\u6b8b\u5dee\u6c60\u5316\u4e5f\u4f7f\u5f97\u7f51\u7edc\u80fd\u591f\u66f4\u597d\u7684\u6355\u6349\u5927\u56fe\u50cf\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002<\/p>\n\n\n\n

RefineNet\u4ee3\u7801\u5b8c\u6574\u5b9e\u73b0\u53ef\u53c2\u8003\uff1a<\/p>\n\n\n\n

https:\/\/github.com\/DrSleep\/refinenet-pytorch<\/a><\/p>\n\n\n\n

\u5176\u4e2d\u5173\u4e8eRCU\u548cCRP\u6a21\u5757\u7684\u5b9e\u73b0\u5982\u4e0b\u4ee3\u7801\u6240\u793a\u3002<\/p>\n\n\n\n

class RCUBlock(nn.Module):\n    \n    def __init__(self, in_planes, out_planes, n_blocks, n_stages):\n        super(RCUBlock, self).__init__()\n        for i in range(n_blocks):\n            for j in range(n_stages):\n                setattr(self, '{}{}'.format(i + 1, stages_suffixes[j]),\n                        conv3x3(in_planes if (i == 0) and (j == 0) else out_planes,\n                                out_planes, stride=1,\n                                bias=(j == 0)))\n        self.stride = 1\n        self.n_blocks = n_blocks\n        self.n_stages = n_stages\n    \n    def forward(self, x):\n        for i in range(self.n_blocks):\n            residual = x\n            for j in range(self.n_stages):\n                x = F.relu(x)\n                x = getattr(self, '{}{}'.format(i + 1, stages_suffixes[j]))(x)\n            x += residual\n        return \n        \nclass CRPBlock(nn.Module):\n\n    def __init__(self, in_planes, out_planes, n_stages):\n        super(CRPBlock, self).__init__()\n        for i in range(n_stages):\n            setattr(self, '{}_{}'.format(i + 1, 'outvar_dimred'),\n                    conv3x3(in_planes if (i == 0) else out_planes,\n                            out_planes, stride=1,\n                            bias=False))\n        self.stride = 1\n        self.n_stages = n_stages\n        self.maxpool = nn.MaxPool2d(kernel_size=5, stride=1, padding=2)\n\n    def forward(self, x):\n        top = x\n        for i in range(self.n_stages):\n            top = self.maxpool(top)\n            top = getattr(self, '{}_{}'.format(i + 1, 'outvar_dimred'))(top)\n            x = top + x\n        return x\n<\/code><\/pre>\n\n\n\n
\u9884\u6d4b\u6548\u679c\uff1a<\/p>\n\n\n\n
\"\"<\/figure><\/div>\n","protected":false},"excerpt":{"rendered":"
\u8bba\u6587\u5730\u5740\uff082016\uff09\uff1aRefineNet: Multi-Path Refinement Networks wi … \u7ee7\u7eed\u9605\u8bfbRefineNet<\/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\/6536"}],"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=6536"}],"version-history":[{"count":5,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/6536\/revisions"}],"predecessor-version":[{"id":6545,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/6536\/revisions\/6545"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=6536"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=6536"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=6536"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}