{"id":225,"date":"2022-03-19T09:54:09","date_gmt":"2022-03-19T09:54:09","guid":{"rendered":"https:\/\/blog.liguanxin.cn\/?p=225"},"modified":"2022-03-19T10:24:16","modified_gmt":"2022-03-19T10:24:16","slug":"%e8%ae%ba%e6%96%87%e7%ac%94%e8%ae%b0-gan-prior-embedded-network-for-blind-face-restoration-in-the-wild","status":"publish","type":"post","link":"https:\/\/blog.liguanxin.cn\/index.php\/2022\/03\/19\/%e8%ae%ba%e6%96%87%e7%ac%94%e8%ae%b0-gan-prior-embedded-network-for-blind-face-restoration-in-the-wild\/","title":{"rendered":"\u8bba\u6587\u7b14\u8bb0\u2014\u2014GAN Prior Embedded Network for Blind Face Restoration in the Wild"},"content":{"rendered":"

\u521b\u65b0\u70b9\uff1a
\n\u2460\u8bad\u7ec3\u4e00\u4e2a\u751f\u6210\u9ad8\u8d28\u91cf\u4eba\u8138\u7684GAN\u7f51\u7edc\uff0c\u7136\u540e\u653e\u5230U\u578b\u7ed3\u6784\u7684\u89e3\u7801\u4fa7\uff0c\u518d\u901a\u8fc7\u4f4e\u5206\u8fa8\u7387\u56fe\u50cf\u8f93\u5165U\u578b\u7ed3\u6784\u4e2d\u8fdb\u884c\u5fae\u8c03\uff08\u4ee5\u524d\u7684\u5de5\u4f5c\u6ca1\u6709\u5fae\u8c03\uff0c\u6307pixel2style2pixel<\/a>\uff09\u3002
\n\u2461\u4ee5\u524d\u7684\u5de5\u4f5c\u6ca1\u6709\u89e3\u51b3\u80cc\u666f\u590d\u6742\u60c5\u51b5\u4e0b\u7684\u95ee\u9898<\/strong><\/p>\n

BFR(Blind face restoration)\u662f\u4e00\u4e2a\u5178\u578b\u7684\u4e0d\u9002\u5b9a\u53cd\u95ee\u9898\uff08ill-posed inverse problem\uff09\u3002
\nwell-posed\uff1a\u89e3\u5b58\u5728\u552f\u4e00\u4ee5\u53ca\u7a33\u5b9a
\n\u53cd\u95ee\u9898\uff1a\u5df2\u77e5y\u6c42x\uff08\u5df2\u77e5\u6a21\u7cca\u56fe\u50cf\u6c42\u539f\u56fe\uff09<\/p>\n

\u52a8\u673a\u548c\u6846\u67b6<\/h2>\n

\u4ee5\u5f80\u7684DNN\u65b9\u6cd5\u662f\u7ed9x\u6c42\u4e00\u4e2ay\uff0c\u800c\u5728\u4e0d\u9002\u5b9a\u95ee\u9898\u4e2dy\u4e0d\u662f\u552f\u4e00\u7684\uff0c\u56e0\u6b64\u4ea7\u751f\u7684y\u901a\u5e38\u662fy\u7a7a\u95f4\u4e2d\u7684\u4e00\u4e2a\u5747\u503c\uff0c\u800c\u8fd9\u4e2a\u503c\u8fc7\u5ea6\u5e73\u6ed1\u800c\u4e14\u4e22\u5931\u5f88\u591a\u7ec6\u8282\u3002
\n\"\"
\n\u672c\u6587\u7684\u65b9\u6cd5\u5c06x\u901a\u8fc7CNN\u7f16\u7801\u5668\u6620\u5c04\u5230\u6f5c\u5728\u7a7a\u95f4\u4e2d\u7684z\uff08\u5bf9\u5e94SytleGAN\u7684\u4ecew\u5230z\uff09\uff0c\u7136\u540eGAN\u7f51\u7edc\u518d\u8f93\u51fa\u56fe\u50cfy\u3002<\/p>\n

\u7ed3\u6784<\/h2>\n

\"\"
\n\u6574\u4f53\u7ed3\u6784\u5c31\u662fCNN\u8ddfStyleGAN2\u7684\u7ed3\u5408\u3002z\u7ecf\u8fc7Mapping\u7f51\u7edc\u4e4b\u540e\u751f\u6210\u7684w\u88ab\u5e7f\u64ad\u5230\u6bcf\u4e2aGAN Block\u4e2d\u3002<\/p>\n

\u635f\u5931\u51fd\u6570<\/h2>\n

\u2460La\u5bf9\u6297\u6027\u635f\u5931\uff08X\u662fground-truth,X~\u662fLQ\u56fe\u50cf\uff09
\n\"\"
\n\u2461Lc\u662f\u5185\u5bb9\u635f\u5931\uff0c\u662f\u751f\u6210\u56fe\u548cground-truth\u4e4b\u95f4\u7684L1\u8303\u6570
\n\u2462Lf\u662f\u611f\u77e5\u635f\u5931\uff0c\u611f\u77e5\u7f51\u7edc\u4e0d\u7528VGG\u800c\u662f\u5224\u522b\u5668
\n\"\"
\n\u6700\u540e\u603b\u635f\u5931<\/strong>
\n\"\"<\/p>\n

CODE<\/h2>\n

\u4e3b\u4f53\u7ed3\u6784<\/p>\n

class FullGenerator(nn.Module):\n    ...\n    def forward(self,\n        inputs,\n        return_latents=False,\n        inject_index=None,\n        truncation=1,\n        truncation_latent=None,\n        input_is_latent=False,\n    ):\n        noise = []\n        # \u9996\u5148\u8fdb\u5165\u5377\u79ef\u5c42\uff0c\u5c42\u6570\u4e3alog(size)\n        for i in range(self.log_size-1):\n            ecd = getattr(self, self.names[i])\n            inputs = ecd(inputs)\n            noise.append(inputs)\n            #print(inputs.shape)\n        inputs = inputs.view(inputs.shape[0], -1)\n        # \u8fdb\u5165FC\u5c42\u63d0\u53d6\u7279\u5f81\n        outs = self.final_linear(inputs)\n        #print(outs.shape)\n        # \u628a\u5217\u8868\u91cd\u590d\u591a\u6b21\uff0c\u4e5f\u5c31\u662f\u5e7f\u64ad\u566a\u58f0\n        noise = list(itertools.chain.from_iterable(itertools.repeat(x, 2) for x in noise))[::-1]\n        # \u8fdb\u5165StyleGAN2\u7f51\u7edc\n        outs = self.generator([outs], return_latents, inject_index, truncation, truncation_latent, input_is_latent, noise=noise[1:])\n        return outs<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"
\u521b\u65b0\u70b9\uff1a \u2460\u8bad\u7ec3\u4e00\u4e2a\u751f\u6210\u9ad8\u8d28\u91cf\u4eba\u8138\u7684GAN\u7f51\u7edc\uff0c\u7136\u540e\u653e\u5230U\u578b\u7ed3\u6784\u7684\u89e3\u7801\u4fa7\uff0c\u518d\u901a\u8fc7\u4f4e\u5206\u8fa8\u7387\u56fe\u50cf\u8f93\u5165U\u578b\u7ed3\u6784\u4e2d\u8fdb\u884c […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[6],"tags":[14,22,20,11,23],"_links":{"self":[{"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/posts\/225"}],"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=225"}],"version-history":[{"count":4,"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/posts\/225\/revisions"}],"predecessor-version":[{"id":299,"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/posts\/225\/revisions\/299"}],"wp:attachment":[{"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/media?parent=225"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/categories?post=225"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.liguanxin.cn\/index.php\/wp-json\/wp\/v2\/tags?post=225"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}