{"id":174,"date":"2022-03-11T03:38:32","date_gmt":"2022-03-11T03:38:32","guid":{"rendered":"https:\/\/blog.liguanxin.cn\/?p=174"},"modified":"2022-03-16T05:27:10","modified_gmt":"2022-03-16T05:27:10","slug":"%e8%ae%ba%e6%96%87%e7%ac%94%e8%ae%b0-fsrnet-end-to-end-learning-face-super-resolution-with-facial-priors","status":"publish","type":"post","link":"https:\/\/blog.liguanxin.cn\/index.php\/2022\/03\/11\/%e8%ae%ba%e6%96%87%e7%ac%94%e8%ae%b0-fsrnet-end-to-end-learning-face-super-resolution-with-facial-priors\/","title":{"rendered":"\u8bba\u6587\u7b14\u8bb0\u2014\u2014FSRNet: End-to-End Learning Face Super-Resolution with Facial Priors"},"content":{"rendered":"

\u521b\u65b0\u70b9\uff1a
\n\u2460\u5229\u7528\u51e0\u4f55\u5148\u9a8c\uff0c\u5373\u9762\u90e8\u5730\u6807\u70ed\u56fe\u548c\u89e3\u6790\u56fe
\n\u2461\u5f15\u5165\u5bf9\u6297\u6027\u7f51\u7edc(FSRGAN)<\/strong><\/p>\n

\u603b\u4f53\u6d41\u7a0b\uff1a\u5148\u7ecf\u8fc7\u4e00\u4e2a\u7f51\u7edc\u6765\u6062\u590d\u7c97\u7cd9\u56fe\u50cf\uff0c\u7136\u540e\u8fdb\u5165\u4e24\u4e2a\u5206\u652f\u5206\u522b\u662f\u7cbe\u7ec6\u7684SR\u7f16\u7801\u5668\u548c\u5148\u9a8c\u4fe1\u606f\u8bc4\u4f30\u7f51\u7edc\u3002\u5148\u9a8c\u4fe1\u606f\u8bc4\u4f30\u7f51\u7edc\u63d0\u53d6\u56fe\u50cf\u7279\u5f81\u7136\u540e\u5bf9landmark\u548cheatmaps\u8fdb\u884c\u8bc4\u4f30\u3002<\/p>\n

\u7f51\u7edc\u7ed3\u6784<\/h3>\n

\"\"<\/p>\n

\u6838\u5fc3\u7ed3\u6784\uff08\u56db\u4e2a\u7f51\u7edc\uff09\uff1a
\nCoarseSRNetwork()
\nFineSREncoder()
\nPriorEstimationNetwork()
\nFineSRDecoder()<\/strong><\/p>\n

\u4ee3\u7801<\/h3>\n
    \n
  • \n

    CoarseSRNetwork<\/p>\n

    class CoarseSRNetwork(nn.Module):\n\ndef __init__(self):\n    super(CoarseSRNetwork, self).__init__()\n    self.conv1 = nn.Sequential(\n        nn.ReflectionPad2d(1),\n        nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=0, bias=False),\n        nn.BatchNorm2d(64),\n        nn.ReLU(True),\n    )\n    self.res_blocks = nn.Sequential(*([ResBlock(64)] * 3))\n    self.conv2 = nn.Sequential(\n        nn.ReflectionPad2d(1),\n        nn.Conv2d(64, 3, kernel_size=3, stride=1, padding=0, bias=False),\n        nn.Tanh(),\n    )\n\ndef forward(self, x):\n    out = self.conv1(x)\n    out = self.res_blocks(out)\n    out = self.conv2(out)\n    return out<\/code><\/pre>\n<\/li>\n
  • \n
    FineSREncoder<\/p>\n
    class FineSREncoder(nn.Module):\n\ndef __init__(self):\n    super(FineSREncoder, self).__init__()\n    self.conv1 = nn.Sequential(\n        nn.Conv2d(3, 64, kernel_size=3, stride=2, padding=1, bias=False),\n        nn.BatchNorm2d(64),\n        nn.ReLU(True),\n    )\n    self.res_blocks = nn.Sequential(*([ResBlock(64)] * 12))\n    self.conv2 = nn.Sequential(\n        nn.ReflectionPad2d(1),\n        nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=0, bias=False),\n        nn.Tanh(),\n    )\n\ndef forward(self, x):\n    out = self.conv1(x)\n    out = self.res_blocks(out)\n    out = self.conv2(out)\n    return out<\/code><\/pre>\n<\/li>\n
  • \n
    PriorEstimationNetwork<\/p>\n
    class PriorEstimationNetwork(nn.Module):\n\ndef __init__(self):\n    super(PriorEstimationNetwork, self).__init__()\n    self.conv1 = nn.Sequential(\n        nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False),\n        nn.BatchNorm2d(64),\n        nn.ReLU(True),\n    )\n    self.res_blocks = nn.Sequential(\n        Residual(64, 128),\n        ResBlock(128),\n        ResBlock(128),\n    )\n    self.hg_blocks = nn.Sequential(\n        HourGlassBlock(128, 3),  # \u6f0f\u6597\u72b6\u7684\u6b8b\u5dee\u5377\u79ef\u7f51\u7edc\n        HourGlassBlock(128, 3),\n    )\n\ndef forward(self, x):\n    out = self.conv1(x)\n    out = self.res_blocks(out)\n    out = self.hg_blocks(out)\n    return out<\/code><\/pre>\n<\/li>\n
  • \n
    FineSRDecoder<\/p>\n
    class FineSRDecoder(nn.Module):\n\ndef __init__(self):\n    super(FineSRDecoder, self).__init__()\n    self.conv1 = nn.Sequential(\n        nn.Conv2d(192, 64, kernel_size=3, stride=1, padding=1, bias=False),\n        nn.BatchNorm2d(64),\n        nn.ReLU(True),\n    )\n    self.deconv1 = nn.Sequential(\n        nn.ConvTranspose2d(64, 64, kernel_size=3, stride=2, padding=1, output_padding=1, bias=False),\n        nn.BatchNorm2d(64),\n        nn.ReLU(True),\n    )\n    self.res_blocks = nn.Sequential(*([ResBlock(64)] * 3))\n    self.conv2 = nn.Sequential(\n        nn.ReflectionPad2d(1),\n        nn.Conv2d(64, 3, kernel_size=3, stride=1, padding=0, bias=False),\n        nn.Tanh(),\n    )\n\ndef forward(self, x):\n    out = self.conv1(x)\n    out = self.deconv1(out)\n    out = self.res_blocks(out)\n    out = self.conv2(out)\n    return out<\/code><\/pre>\n<\/li>\n
  • \n
    \u4e3b\u4f53\u7ed3\u6784<\/p>\n
    class FSRNet(nn.Module):\n\ndef __init__(self, hmaps_ch, pmaps_ch):\n    ...\n\ndef forward(self, x):\n    y_c = self.csr_net(x)  # \u7c97\u7cd9\u7f51\u7edc\n    f = self.fsr_enc(y_c)  # \u7cbe\u7ec6\u7684SR\u7f16\u7801\u5668\n    p = self.pre_net(y_c)  # \u5148\u9a8c\u7f51\u7edc\n\n    # 1x1 conv for hmaps & pmaps\uff08\u6784\u9020\u5148\u9a8c\u8bc4\u4f30\u7684\u7ed3\u679c\uff09\n    b1 = (self.prior_conv1 is not None)\n    b2 = (self.prior_conv2 is not None)\n    if b1 and b2:\n        hmaps = self.prior_conv1(p)\n        pmaps = self.prior_conv2(p)\n        prs = torch.cat((hmaps, pmaps), 1)\n    elif b1:\n        prs = self.prior_conv1(p)\n    elif b2:\n        prs = self.prior_conv2(p)\n\n    concat = torch.cat((f, p), 1)  # \u5408\u5e76sr\u7f16\u7801\u5668\n    out = self.fsr_dec(concat)  # sr\u89e3\u7801\u5668\n    return y_c, prs, out<\/code><\/pre>\n<\/li>\n
  • \n
    \u635f\u5931\u8ba1\u7b97<\/p>\n
        loss1 = criterion(y_c, image_hr)\n    loss2 = criterion(out, image_hr)\n    loss3 = criterion(prs, image_pr)\n    loss = loss1 + loss2 + loss3<\/code><\/pre>\n<\/li>\n<\/ul>\n
    \u7591\u95ee<\/h3>\n
    \u4e3a\u4f55\u5377\u79ef\u6784\u6210\u7684PriorEstimationNetwork\u80fd\u63d0\u53d6\u5148\u9a8c\u4fe1\u606f\uff1f
    \n\u7b54\uff1a\u56e0\u4e3a\u8fd9\u4e2a\u7f51\u7edc\u7684\u635f\u5931\u662f\u5355\u72ec\u8ba1\u7b97\u7684\uff0c\u4e0egroundtruth\u7684hmaps\u548cpmaps\u6765\u6bd4\u8f83\u3002
    \n\"\"<\/p>\n","protected":false},"excerpt":{"rendered":"
    \u521b\u65b0\u70b9\uff1a \u2460\u5229\u7528\u51e0\u4f55\u5148\u9a8c\uff0c\u5373\u9762\u90e8\u5730\u6807\u70ed\u56fe\u548c\u89e3\u6790\u56fe \u2461\u5f15\u5165\u5bf9\u6297\u6027\u7f51\u7edc(FSRGAN) \u603b\u4f53\u6d41\u7a0b\uff1a\u5148\u7ecf\u8fc7\u4e00\u4e2a\u7f51\u7edc\u6765 […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[6],"tags":[14,20,11],"_links":{"self":[{"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/posts\/174"}],"collection":[{"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/comments?post=174"}],"version-history":[{"count":3,"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/posts\/174\/revisions"}],"predecessor-version":[{"id":180,"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/posts\/174\/revisions\/180"}],"wp:attachment":[{"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/media?parent=174"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/categories?post=174"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/tags?post=174"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}