{"id":859,"date":"2025-07-28T01:34:28","date_gmt":"2025-07-27T17:34:28","guid":{"rendered":"https:\/\/blog.sssn.tech\/?p=859"},"modified":"2025-07-28T01:41:35","modified_gmt":"2025-07-27T17:41:35","slug":"swinir-image-restoration-using-swin-transformer-%e5%85%a8%e6%96%87%e7%bf%bb%e8%af%91%e6%b3%a8%e9%87%8a","status":"publish","type":"post","link":"https:\/\/blog.sssn.tech\/?p=859","title":{"rendered":"SwinIR: Image Restoration Using Swin Transformer \u5168\u6587\u7ffb\u8bd1+\u6ce8\u91ca"},"content":{"rendered":"<h1>SwinIR: Image Restoration Using Swin Transformer<\/h1>\n<h2>Abstract<\/h2>\n<p>\u56fe\u50cf\u590d\u539f\u662f\u4e00\u4e2a\u7531\u6765\u5df2\u4e45\u7684\u4f4e\u5c42\u89c6\u89c9\u95ee\u9898\uff0c\u5176\u76ee\u6807\u662f\u4ece\u4f4e\u8d28\u91cf\u56fe\u50cf\uff08\u4f8b\u5982\u964d\u91c7\u6837\u3001\u5e26\u566a\u58f0\u6216\u538b\u7f29\u56fe\u50cf\uff09\u4e2d\u8fd8\u539f\u51fa\u9ad8\u8d28\u91cf\u56fe\u50cf\u3002\u5c3d\u7ba1\u5f53\u524d\u6700\u5148\u8fdb\u7684\u56fe\u50cf\u590d\u539f\u65b9\u6cd5\u5927\u591a\u57fa\u4e8e\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff0c\u4f46\u5728\u9ad8\u5c42\u89c6\u89c9\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u8272\u7684Transformer\u5c1a\u5c11\u6709\u5c1d\u8bd5\u5e94\u7528\u4e8e\u6b64\u9886\u57df\u3002<\/p>\n<p>\u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u4e2a\u57fa\u4e8eSwin Transformer\u7684\u5f3a\u57fa\u7ebf\u6a21\u578b\u2014\u2014SwinIR\uff0c\u7528\u4e8e\u56fe\u50cf\u590d\u539f\u4efb\u52a1\u3002SwinIR\u4e3b\u8981\u7531\u4e09\u4e2a\u90e8\u5206\u7ec4\u6210\uff1a\u6d45\u5c42\u7279\u5f81\u63d0\u53d6<strong>\u3001<\/strong>\u6df1\u5c42\u7279\u5f81\u63d0\u53d6\u548c\u9ad8\u8d28\u91cf\u56fe\u50cf\u91cd\u5efa\u3002\u5176\u4e2d\uff0c\u6df1\u5c42\u7279\u5f81\u63d0\u53d6\u6a21\u5757\u7531\u591a\u4e2a\u6b8b\u5deeSwin Transformer\u5757\uff08Residual Swin Transformer Block\uff0cRSTB\uff09\u6784\u6210\uff0c\u6bcf\u4e2a\u5757\u5185\u90e8\u5305\u542b\u591a\u4e2aSwin Transformer\u5c42\uff0c\u5e76\u914d\u6709\u6b8b\u5dee\u8fde\u63a5\u7ed3\u6784\u3002<\/p>\n<p>\u6211\u4eec\u5728\u4e09\u4e2a\u5177\u6709\u4ee3\u8868\u6027\u7684\u4efb\u52a1\u4e0a\u8fdb\u884c\u4e86\u5b9e\u9a8c\uff1a\u56fe\u50cf\u8d85\u5206\u8fa8\u7387\uff08\u5305\u62ec\u7ecf\u5178\u3001\u8f7b\u91cf\u7ea7\u548c\u771f\u5b9e\u573a\u666f\u56fe\u50cf\u8d85\u5206\u8fa8\u7387\uff09\u3001\u56fe\u50cf\u53bb\u566a\uff08\u5305\u62ec\u7070\u5ea6\u56fe\u50cf\u548c\u5f69\u8272\u56fe\u50cf\u53bb\u566a\uff09\u4ee5\u53caJPEG\u538b\u7f29\u4f2a\u5f71\u53bb\u9664\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cSwinIR\u5728\u591a\u4e2a\u4efb\u52a1\u4e2d\u5747\u4f18\u4e8e\u73b0\u6709\u6700\u5148\u8fdb\u7684\u65b9\u6cd5\uff0c\u6027\u80fd\u63d0\u5347\u5e45\u5ea6\u53ef\u8fbe0.14\u223c0.45dB\uff0c\u540c\u65f6\u6a21\u578b\u53c2\u6570\u603b\u91cf\u6700\u591a\u53ef\u51cf\u5c1167%\u3002<\/p>\n<h2>1. Introduction<\/h2>\n<p>\u56fe\u50cf\u590d\u539f\u4efb\u52a1\uff08\u5982\u56fe\u50cf\u8d85\u5206\u8fa8\u7387\uff08SR\uff09\u3001\u56fe\u50cf\u53bb\u566a\u548cJPEG\u538b\u7f29\u4f2a\u5f71\u53bb\u9664\uff09\u65e8\u5728\u4ece\u4f4e\u8d28\u91cf\u9000\u5316\u56fe\u50cf\u4e2d\u91cd\u5efa\u51fa\u9ad8\u8d28\u91cf\u7684\u5e72\u51c0\u56fe\u50cf\u3002\u81ea\u4ece\u4e00\u7cfb\u5217\u5177\u6709\u9769\u547d\u6027\u5f71\u54cd\u7684\u5de5\u4f5c\u95ee\u4e16\u4ee5\u6765 [18, 40, 90, 91]\uff0c\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08CNN\uff09\u5df2\u7ecf\u6210\u4e3a\u56fe\u50cf\u590d\u539f\u7684\u4e3b\u529b\u65b9\u6cd5 [43, 51, 43, 81, 92, 95, 24, 93, 46, 89, 88]\u3002\u5927\u591a\u6570\u57fa\u4e8eCNN\u7684\u65b9\u6cd5\u4e3b\u8981\u805a\u7126\u4e8e\u7cbe\u5fc3\u8bbe\u8ba1\u7684\u7f51\u7edc\u67b6\u6784\uff0c\u5982\u6b8b\u5dee\u5b66\u4e60 [43, 51] \u548c\u5bc6\u96c6\u8fde\u63a5 [97, 81]\u3002\u867d\u7136\u4e0e\u4f20\u7edf\u7684\u57fa\u4e8e\u6a21\u578b\u7684\u65b9\u6cd5\u76f8\u6bd4\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u5728\u6027\u80fd\u4e0a\u6709\u4e86\u663e\u8457\u63d0\u5347 [73, 14, 28]\uff0c\u4f46\u5b83\u4eec\u666e\u904d\u5b58\u5728\u4e24\u4e2a\u6e90\u81ea\u5377\u79ef\u5c42\u672c\u8eab\u7684\u57fa\u672c\u95ee\u9898\uff1a<\/p>\n<ol>\n<li>\u56fe\u50cf\u4e0e\u5377\u79ef\u6838\u4e4b\u95f4\u7684\u4ea4\u4e92\u662f\u5185\u5bb9\u65e0\u5173\u7684\u3002\u5bf9\u4e0d\u540c\u56fe\u50cf\u533a\u57df\u4f7f\u7528\u76f8\u540c\u7684\u5377\u79ef\u6838\u53ef\u80fd\u5e76\u975e\u6700\u4f18\u9009\u62e9\u3002<\/li>\n<li>\u7531\u4e8e\u5377\u79ef\u9075\u5faa\u5c40\u90e8\u5904\u7406\u539f\u5219\uff0c\u5b83\u5728\u5efa\u6a21\u957f\u8ddd\u79bb\u4f9d\u8d56\u65f6\u6548\u679c\u6709\u9650\u3002<\/li>\n<\/ol>\n<p>\u4f5c\u4e3aCNN\u7684\u66ff\u4ee3\u65b9\u6848\uff0cTransformer [76] \u8bbe\u8ba1\u4e86\u4e00\u79cd\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff0c\u7528\u4e8e\u6355\u6349\u5168\u5c40\u4e0a\u4e0b\u6587\u4e4b\u95f4\u7684\u4ea4\u4e92\uff0c\u5e76\u5728\u591a\u4e2a\u89c6\u89c9\u4efb\u52a1\u4e2d\u5c55\u73b0\u51fa\u826f\u597d\u7684\u6027\u80fd [6, 74, 19, 56]\u3002\u7136\u800c\uff0c\u5e94\u7528\u4e8e\u56fe\u50cf\u590d\u539f\u7684\u89c6\u89c9Transformer\u65b9\u6cd5 [9, 5] \u901a\u5e38\u5c06\u8f93\u5165\u56fe\u50cf\u5212\u5206\u4e3a\u56fa\u5b9a\u5927\u5c0f\u7684\u56fe\u50cf\u5757\uff08\u598248\u00d748\uff09\uff0c\u5e76\u72ec\u7acb\u5904\u7406\u6bcf\u4e2a\u56fe\u5757\u3002\u8fd9\u79cd\u7b56\u7565\u4e0d\u53ef\u907f\u514d\u5730\u5e26\u6765\u4e24\u4e2a\u7f3a\u9677\uff1a<\/p>\n<ol>\n<li>\u56fe\u5757\u8fb9\u7f18\u50cf\u7d20\u65e0\u6cd5\u5229\u7528\u56fe\u5757\u4ee5\u5916\u7684\u90bb\u8fd1\u50cf\u7d20\u8fdb\u884c\u56fe\u50cf\u590d\u539f\uff1b<\/li>\n<li>\u91cd\u5efa\u540e\u7684\u56fe\u50cf\u53ef\u80fd\u5728\u6bcf\u4e2a\u56fe\u5757\u8fb9\u7f18\u51fa\u73b0\u4f2a\u5f71\u3002<\/li>\n<\/ol>\n<p>\u867d\u7136\u901a\u8fc7\u56fe\u5757\u91cd\u53e0\u53ef\u4ee5\u7f13\u89e3\u8fd9\u4e00\u95ee\u9898\uff0c\u4f46\u8fd9\u4f1a\u663e\u8457\u589e\u52a0\u8ba1\u7b97\u8d1f\u62c5\u3002<\/p>\n<p>\u8fd1\u5e74\u6765\uff0cSwin Transformer [56] \u5c55\u73b0\u51fa\u6781\u5927\u7684\u6f5c\u529b\uff0c\u5b83\u7ed3\u5408\u4e86CNN\u4e0eTransformer\u7684\u4f18\u70b9\uff1a<\/p>\n<ul>\n<li>\u4e00\u65b9\u9762\uff0c\u501f\u52a9\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236\uff0c\u5b83\u5177\u5907CNN\u5728\u5904\u7406\u5927\u5c3a\u5bf8\u56fe\u50cf\u65b9\u9762\u7684\u4f18\u52bf\uff1b<\/li>\n<li>\u53e6\u4e00\u65b9\u9762\uff0c\u901a\u8fc7\u6ed1\u52a8\u7a97\u53e3\u673a\u5236\uff08shifted window scheme\uff09\uff0c\u5b83\u80fd\u591f\u6709\u6548\u5efa\u6a21\u957f\u8ddd\u79bb\u4f9d\u8d56\u3002<\/li>\n<\/ul>\n<p>\u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u57fa\u4e8eSwin Transformer\u63d0\u51fa\u4e86\u4e00\u79cd\u56fe\u50cf\u590d\u539f\u6a21\u578b\uff0c\u79f0\u4e3aSwinIR\u3002\u5177\u4f53\u6765\u8bf4\uff0cSwinIR\u5305\u542b\u4e09\u4e2a\u6a21\u5757\uff1a<\/p>\n<ol>\n<li>\u6d45\u5c42\u7279\u5f81\u63d0\u53d6\u6a21\u5757\uff1a\u4f7f\u7528\u4e00\u4e2a\u5377\u79ef\u5c42\u63d0\u53d6\u6d45\u5c42\u7279\u5f81\uff0c\u5e76\u5c06\u5176\u76f4\u63a5\u4f20\u9012\u7ed9\u91cd\u5efa\u6a21\u5757\uff0c\u4ee5\u4fdd\u7559\u4f4e\u9891\u4fe1\u606f\uff1b<\/li>\n<li>\u6df1\u5c42\u7279\u5f81\u63d0\u53d6\u6a21\u5757\uff1a\u4e3b\u8981\u7531\u591a\u4e2a\u6b8b\u5deeSwin Transformer\u5757\uff08RSTB\uff09\u7ec4\u6210\uff0c\u6bcf\u4e2a\u5757\u5305\u542b\u591a\u4e2aSwin Transformer\u5c42\uff0c\u7528\u4e8e\u5c40\u90e8\u6ce8\u610f\u529b\u548c\u8de8\u7a97\u53e3\u4ea4\u4e92\u3002\u6b64\u5916\uff0c\u6211\u4eec\u5728\u6bcf\u4e2a\u5757\u672b\u5c3e\u52a0\u5165\u4e00\u4e2a\u5377\u79ef\u5c42\u4ee5\u589e\u5f3a\u7279\u5f81\uff0c\u5e76\u4f7f\u7528\u6b8b\u5dee\u8fde\u63a5\u4ee5\u5b9e\u73b0\u7279\u5f81\u805a\u5408\u7684\u6377\u5f84\uff1b<\/li>\n<li>\u9ad8\u8d28\u91cf\u56fe\u50cf\u91cd\u5efa\u6a21\u5757\uff1a\u5c06\u6d45\u5c42\u4e0e\u6df1\u5c42\u7279\u5f81\u878d\u5408\uff0c\u7528\u4e8e\u6700\u7ec8\u7684\u9ad8\u8d28\u91cf\u56fe\u50cf\u91cd\u5efa\u3002<\/li>\n<\/ol>\n<p>\u4e0e\u4e3b\u6d41\u7684CNN\u56fe\u50cf\u590d\u539f\u6a21\u578b\u76f8\u6bd4\uff0c\u57fa\u4e8eTransformer\u7684SwinIR\u5177\u6709\u4ee5\u4e0b\u51e0\u4e2a\u4f18\u52bf\uff1a<\/p>\n<ol>\n<li>\u57fa\u4e8e\u5185\u5bb9\u7684\u4ea4\u4e92\u673a\u5236\uff1a\u56fe\u50cf\u5185\u5bb9\u4e0e\u6ce8\u610f\u529b\u6743\u91cd\u7684\u4ea4\u4e92\u53ef\u4ee5\u88ab\u89e3\u91ca\u4e3a\u4e00\u79cd\u7a7a\u95f4\u53d8\u5316\u7684\u5377\u79ef [13, 21, 75]\uff1b<\/li>\n<li>\u957f\u8ddd\u79bb\u4f9d\u8d56\u5efa\u6a21\u80fd\u529b\uff1a\u5f97\u76ca\u4e8e\u6ed1\u52a8\u7a97\u53e3\u673a\u5236\uff1b<\/li>\n<li>\u5728\u66f4\u5c11\u53c2\u6570\u4e0b\u5b9e\u73b0\u66f4\u4f18\u6027\u80fd\uff1a\u4f8b\u5982\uff0c\u5982\u56fe1\u6240\u793a\uff0cSwinIR\u5728\u53c2\u6570\u66f4\u5c11\u7684\u60c5\u51b5\u4e0b\uff0c\u76f8\u6bd4\u73b0\u6709\u56fe\u50cf\u8d85\u5206\u8fa8\u7387\u65b9\u6cd5\uff0c\u80fd\u83b7\u5f97\u66f4\u9ad8\u7684PSNR\uff08\u5cf0\u503c\u4fe1\u566a\u6bd4\uff09\u6027\u80fd\u3002<\/li>\n<\/ol>\n<p><img decoding=\"async\" src=\"https:\/\/pic.cirno.fun\/sssn-blog-pics\/image-20250727180556317.png\" alt=\"image-20250727180556317\" style=\"zoom:66%;\" \/><\/p>\n<h2>2. Related Work<\/h2>\n<h3>2.1. Image Restoration<\/h3>\n<p>\u4e0e\u4f20\u7edf\u7684\u56fe\u50cf\u590d\u539f\u65b9\u6cd5 [28, 72, 73, 62, 32]\uff08\u901a\u5e38\u662f\u57fa\u4e8e\u6a21\u578b\u7684\u65b9\u6cd5\uff09\u76f8\u6bd4\uff0c\u57fa\u4e8e\u5b66\u4e60\u7684\u65b9\u6cd5\uff0c\u5c24\u5176\u662f\u57fa\u4e8eCNN\u7684\u65b9\u6cd5\uff0c\u7531\u4e8e\u5176\u51fa\u8272\u7684\u6027\u80fd\uff0c\u5df2\u53d8\u5f97\u66f4\u52a0\u6d41\u884c\u3002\u8fd9\u7c7b\u65b9\u6cd5\u901a\u5e38\u901a\u8fc7\u5927\u89c4\u6a21\u6210\u5bf9\u7684\u6570\u636e\u96c6\uff0c\u5b66\u4e60\u4f4e\u8d28\u91cf\u56fe\u50cf\u4e0e\u9ad8\u8d28\u91cf\u56fe\u50cf\u4e4b\u95f4\u7684\u6620\u5c04\u5173\u7cfb\u3002<\/p>\n<p>\u81ea\u4ece\u5f00\u521b\u6027\u7684\u5de5\u4f5c\u2014\u2014SRCNN[18]\uff08\u7528\u4e8e\u56fe\u50cf\u8d85\u5206\u8fa8\u7387\uff09\u3001DnCNN [90]\uff08\u7528\u4e8e\u56fe\u50cf\u53bb\u566a\uff09\u4ee5\u53caARCNN [17]\uff08\u7528\u4e8eJPEG\u538b\u7f29\u4f2a\u5f71\u53bb\u9664\uff09\u95ee\u4e16\u4ee5\u6765\uff0c\u6d8c\u73b0\u51fa\u5927\u91cf\u57fa\u4e8eCNN\u7684\u6a21\u578b\uff0c\u65e8\u5728\u901a\u8fc7\u66f4\u590d\u6742\u7684\u795e\u7ecf\u7f51\u7edc\u7ed3\u6784\u8bbe\u8ba1\u63d0\u5347\u6a21\u578b\u7684\u8868\u793a\u80fd\u529b\u3002<\/p>\n<p>\u8fd9\u4e9b\u7ed3\u6784\u8bbe\u8ba1\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\u6b8b\u5dee\u5757\uff08Residual Block\uff09[40, 7, 88]\uff1b<\/li>\n<li>\u5bc6\u96c6\u5757\uff08Dense Block\uff09[81, 97, 98]\uff1b<\/li>\n<li>\u4ee5\u53ca\u5176\u4ed6\u591a\u79cd\u7ed3\u6784 [10, 42, 93, 78, 77, 79, 50, 48, 49, 92, 70, 36, 83, 30, 11, 16, 96, 64, 38, 26, 41, 25]\u3002<\/li>\n<\/ul>\n<p>\u6b64\u5916\uff0c\u4e00\u4e9b\u65b9\u6cd5\u8fd8\u5728CNN\u6846\u67b6\u4e2d\u5f15\u5165\u4e86\u6ce8\u610f\u529b\u673a\u5236\uff0c\u4f8b\u5982\uff1a<\/p>\n<ul>\n<li>\u901a\u9053\u6ce8\u610f\u529b\uff08Channel Attention\uff09[95, 15, 63]\uff1b<\/li>\n<li>\u975e\u5c40\u90e8\u6ce8\u610f\u529b\uff08Non-local Attention\uff09[52, 61]\uff1b<\/li>\n<li>\u81ea\u9002\u5e94\u56fe\u5757\u805a\u5408\uff08Adaptive Patch Aggregation\uff09[100]\u3002<\/li>\n<\/ul>\n<h3>2.2 Vision Transformer<\/h3>\n<p>\u8fd1\u5e74\u6765\uff0c\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4e2d\u7684\u6a21\u578b Transformer [76] \u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u53d8\u5f97\u8d8a\u6765\u8d8a\u53d7\u6b22\u8fce\u3002\u5f53\u5e94\u7528\u4e8e\u56fe\u50cf\u5206\u7c7b [66, 19, 84, 56, 45, 55, 75]\u3001\u76ee\u6807\u68c0\u6d4b [6, 53, 74, 56]\u3001\u56fe\u50cf\u5206\u5272 [84, 99, 56, 4] \u548c\u4eba\u7fa4\u8ba1\u6570 [47, 69] \u7b49\u89c6\u89c9\u4efb\u52a1\u65f6\uff0cTransformer \u901a\u8fc7\u63a2\u7d22\u56fe\u50cf\u4e0d\u540c\u533a\u57df\u4e4b\u95f4\u7684\u5168\u5c40\u4ea4\u4e92\uff0c\u5b66\u4e60\u5173\u6ce8\u5173\u952e\u56fe\u50cf\u533a\u57df\u3002\u5f97\u76ca\u4e8e\u5176\u51fa\u8272\u7684\u8868\u73b0\uff0cTransformer \u4e5f\u88ab\u5f15\u5165\u5230\u56fe\u50cf\u590d\u539f\u4efb\u52a1\u4e2d [9, 5, 82]\u3002Chen \u7b49\u4eba [9] \u57fa\u4e8e\u6807\u51c6 Transformer \u63d0\u51fa\u4e86\u4e00\u4e2a\u4e3b\u5e72\u6a21\u578b IPT\uff0c\u7528\u4e8e\u89e3\u51b3\u591a\u79cd\u56fe\u50cf\u590d\u539f\u95ee\u9898\u3002\u7136\u800c\uff0cIPT \u7684\u6027\u80fd\u4f9d\u8d56\u4e8e\u5927\u91cf\u53c2\u6570\uff08\u8d85\u8fc7 1.155 \u4ebf\u4e2a\u53c2\u6570\uff09\u3001\u5927\u89c4\u6a21\u6570\u636e\u96c6\uff08\u8d85\u8fc7 110 \u4e07\u5f20\u56fe\u50cf\uff09\u4ee5\u53ca\u591a\u4efb\u52a1\u5b66\u4e60\u7b56\u7565\u3002Cao \u7b49\u4eba [5] \u63d0\u51fa\u4e86 VSR-Transformer\uff0c\u8be5\u6a21\u578b\u5728\u89c6\u9891\u8d85\u5206\u8fa8\u7387\u4efb\u52a1\u4e2d\u5229\u7528\u81ea\u6ce8\u610f\u529b\u673a\u5236\u5b9e\u73b0\u66f4\u4f18\u7684\u7279\u5f81\u878d\u5408\uff0c\u4f46\u5176\u56fe\u50cf\u7279\u5f81\u4ecd\u7136\u901a\u8fc7 CNN \u63d0\u53d6\u3002\u6b64\u5916\uff0cIPT \u548c VSR-Transformer \u90fd\u91c7\u7528\u57fa\u4e8e\u56fe\u50cf\u5757\u7684\u6ce8\u610f\u529b\u673a\u5236\uff0c\u8fd9\u53ef\u80fd\u5e76\u4e0d\u9002\u7528\u4e8e\u56fe\u50cf\u590d\u539f\u4efb\u52a1\u3002\u53e6\u5916\uff0c\u4e00\u9879\u540c\u671f\u7684\u5de5\u4f5c [82] \u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e Swin Transformer [56] \u7684 U \u5f62\u7ed3\u6784\u3002<\/p>\n<h2>3. Method<\/h2>\n<h3>3.1 Network Architechture<\/h3>\n<p><img decoding=\"async\" src=\"https:\/\/pic.cirno.fun\/sssn-blog-pics\/image-20250727190548339.png\" alt=\"image-20250727190548339\" style=\"zoom:33%;\" \/><\/p>\n<p>\u5982\u56fe 2 \u6240\u793a\uff0cSwinIR \u5305\u542b\u4e09\u4e2a\u6a21\u5757\uff1a\u6d45\u5c42\u7279\u5f81\u63d0\u53d6\u3001\u6df1\u5c42\u7279\u5f81\u63d0\u53d6\u548c\u9ad8\u8d28\u91cf\uff08HQ\uff09\u56fe\u50cf\u91cd\u5efa\u6a21\u5757\u3002\u6211\u4eec\u5728\u6240\u6709\u56fe\u50cf\u590d\u539f\u4efb\u52a1\u4e2d\u4f7f\u7528\u76f8\u540c\u7684\u7279\u5f81\u63d0\u53d6\u6a21\u5757\uff0c\u4f46\u9488\u5bf9\u4e0d\u540c\u4efb\u52a1\u4f7f\u7528\u4e0d\u540c\u7684\u91cd\u5efa\u6a21\u5757\u3002<\/p>\n<h4>\u6d45\u5c42\u4e0e\u6df1\u5c42\u7279\u5f81\u63d0\u53d6<\/h4>\n<p>\u7ed9\u5b9a\u4e00\u4e2a\u4f4e\u8d28\u91cf\uff08LQ\uff09\u8f93\u5165\u56fe\u50cf <code class=\"katex-inline\">I_{LQ} \\in \\mathbb{R}^{H \\times W \\times C_{in}}<\/code>\uff08\u5176\u4e2d <code class=\"katex-inline\">H<\/code>\u3001<code class=\"katex-inline\">W<\/code> \u548c <code class=\"katex-inline\">C_{in}<\/code> \u5206\u522b\u8868\u793a\u56fe\u50cf\u7684\u9ad8\u5ea6\u3001\u5bbd\u5ea6\u548c\u8f93\u5165\u901a\u9053\u6570\uff09\uff0c\u6211\u4eec\u4f7f\u7528\u4e00\u4e2a <code class=\"katex-inline\">3 \\times 3<\/code> \u7684\u5377\u79ef\u5c42 <code class=\"katex-inline\">H_{SF}(\\cdot)<\/code> \u6765\u63d0\u53d6\u6d45\u5c42\u7279\u5f81 <code class=\"katex-inline\">F_0 \\in \\mathbb{R}^{H \\times W \\times C}<\/code>\uff0c\u8ba1\u7b97\u65b9\u5f0f\u5982\u4e0b\uff1a<\/p>\n<pre><code class=\"language-katex\">F_0 = H_{SF}(I_{LQ})<\/code><\/pre>\n<p>\u5176\u4e2d <code class=\"katex-inline\">C<\/code> \u8868\u793a\u7279\u5f81\u901a\u9053\u6570\u3002\u8be5\u5377\u79ef\u5c42\u5728\u56fe\u50cf\u5904\u7406\u521d\u671f\u8868\u73b0\u826f\u597d\uff0c\u6709\u52a9\u4e8e\u66f4\u7a33\u5b9a\u7684\u4f18\u5316\u548c\u53d6\u5f97\u66f4\u4f18\u7684\u7ed3\u679c [86]\u3002\u5b83\u8fd8\u63d0\u4f9b\u4e86\u4e00\u79cd\u7b80\u5355\u7684\u65b9\u6cd5\uff0c\u5c06\u8f93\u5165\u56fe\u50cf\u7a7a\u95f4\u6620\u5c04\u5230\u66f4\u9ad8\u7ef4\u7684\u7279\u5f81\u7a7a\u95f4\u3002<\/p>\n<p>\u63a5\u7740\uff0c\u6211\u4eec\u4ece <code class=\"katex-inline\">F_0<\/code> \u4e2d\u63d0\u53d6\u6df1\u5c42\u7279\u5f81 <code class=\"katex-inline\">F_{DF} \\in \\mathbb{R}^{H \\times W \\times C}<\/code>\uff0c\u5176\u8ba1\u7b97\u5982\u4e0b\uff1a<\/p>\n<pre><code class=\"language-katex\">F_{DF} = H_{DF}(F_0)<\/code><\/pre>\n<p>\u5176\u4e2d <code class=\"katex-inline\">H_{DF}(\\cdot)<\/code> \u662f\u6df1\u5c42\u7279\u5f81\u63d0\u53d6\u6a21\u5757\uff0c\u5b83\u5305\u542b <code class=\"katex-inline\">K<\/code> \u4e2a\u6b8b\u5dee Swin Transformer \u5757\uff08RSTB\uff09\u548c\u4e00\u4e2a <code class=\"katex-inline\">3 \\times 3<\/code> \u5377\u79ef\u5c42\u3002\u66f4\u5177\u4f53\u5730\u8bf4\uff0c\u4e2d\u95f4\u7279\u5f81 <code class=\"katex-inline\">F_1, F_2, \\ldots, F_K<\/code> \u4ee5\u53ca\u6700\u7ec8\u7684\u6df1\u5c42\u7279\u5f81 <code class=\"katex-inline\">F_{DF}<\/code> \u662f\u9010\u5757\u63d0\u53d6\u7684\uff0c\u8ba1\u7b97\u5982\u4e0b\uff1a<\/p>\n<pre><code class=\"language-katex\">F_i = H_{RSTB_i}(F_{i-1}), \\quad i = 1, 2, \\ldots, K<\/code><\/pre>\n<pre><code class=\"language-katex\">F_{DF} = H_{CONV}(F_K)<\/code><\/pre>\n<p>\u5176\u4e2d <code class=\"katex-inline\">H_{RSTB_i}(\\cdot)<\/code> \u8868\u793a\u7b2c <code class=\"katex-inline\">i<\/code> \u4e2a RSTB\uff08\u6b8b\u5dee Swin Transformer \u5757\uff09\uff0c<code class=\"katex-inline\">H_{CONV}<\/code> \u662f\u6700\u540e\u7684\u5377\u79ef\u5c42\u3002\u5728\u7279\u5f81\u63d0\u53d6\u672b\u5c3e\u4f7f\u7528\u5377\u79ef\u5c42\uff0c\u53ef\u4ee5\u5c06\u5377\u79ef\u64cd\u4f5c\u7684\u5f52\u7eb3\u504f\u7f6e\u5f15\u5165\u5230\u57fa\u4e8e Transformer \u7684\u7f51\u7edc\u4e2d\uff0c\u4e3a\u540e\u7eed\u7684\u6d45\u5c42\u4e0e\u6df1\u5c42\u7279\u5f81\u878d\u5408\u6253\u4e0b\u66f4\u597d\u7684\u57fa\u7840\u3002<\/p>\n<h4>Image reconstruction<\/h4>\n<p>\u4ee5\u56fe\u50cf\u8d85\u5206\u8fa8\uff08SR\uff09\u4e3a\u4f8b\uff0c\u6211\u4eec\u901a\u8fc7\u805a\u5408\u6d45\u5c42\u548c\u6df1\u5c42\u7279\u5f81\u6765\u91cd\u5efa\u9ad8\u8d28\u91cf\u56fe\u50cf <code class=\"katex-inline\">I_{RHQ}<\/code>\uff0c\u8ba1\u7b97\u65b9\u5f0f\u5982\u4e0b\uff1a<\/p>\n<pre><code class=\"language-katex\">I_{RHQ} = H_{REC}(F_0 + F_{DF})<\/code><\/pre>\n<p>\u5176\u4e2d <code class=\"katex-inline\">H_{REC}(\\cdot)<\/code> \u662f\u91cd\u5efa\u6a21\u5757\u7684\u51fd\u6570\u3002\u6d45\u5c42\u7279\u5f81\u4e3b\u8981\u5305\u542b\u4f4e\u9891\u4fe1\u606f\uff0c\u800c\u6df1\u5c42\u7279\u5f81\u5219\u4e13\u6ce8\u4e8e\u6062\u590d\u4e22\u5931\u7684\u9ad8\u9891\u4fe1\u606f\u3002\u501f\u52a9\u957f\u8df3\u8dc3\u8fde\u63a5\uff0cSwinIR \u53ef\u4ee5\u5c06\u4f4e\u9891\u4fe1\u606f\u76f4\u63a5\u4f20\u9012\u7ed9\u91cd\u5efa\u6a21\u5757\uff0c\u6709\u52a9\u4e8e\u6df1\u5c42\u7279\u5f81\u63d0\u53d6\u6a21\u5757\u4e13\u6ce8\u4e8e\u9ad8\u9891\u4fe1\u606f\u5e76\u7a33\u5b9a\u8bad\u7ec3\u3002<\/p>\n<p>\u5728\u91cd\u5efa\u6a21\u5757\u7684\u5b9e\u73b0\u4e2d\uff0c\u6211\u4eec\u4f7f\u7528\u4e9a\u50cf\u7d20\u5377\u79ef\u5c42\uff08sub-pixel convolution layer\uff09[68] \u6765\u5bf9\u7279\u5f81\u8fdb\u884c\u4e0a\u91c7\u6837\u3002<\/p>\n<p>\u5bf9\u4e8e\u4e0d\u9700\u8981\u4e0a\u91c7\u6837\u7684\u4efb\u52a1\uff0c\u5982\u56fe\u50cf\u53bb\u566a\u548c JPEG \u538b\u7f29\u4f2a\u5f71\u53bb\u9664\uff0c\u4ec5\u4f7f\u7528\u4e00\u4e2a\u5377\u79ef\u5c42\u8fdb\u884c\u91cd\u5efa\u3002\u6b64\u5916\uff0c\u6211\u4eec\u91c7\u7528\u6b8b\u5dee\u5b66\u4e60\u6765\u91cd\u5efa LQ \u56fe\u50cf\u4e0e HQ \u56fe\u50cf\u4e4b\u95f4\u7684\u6b8b\u5dee\uff0c\u800c\u4e0d\u662f\u76f4\u63a5\u91cd\u5efa HQ \u56fe\u50cf\u3002\u5176\u516c\u5f0f\u5982\u4e0b\uff1a<\/p>\n<pre><code class=\"language-katex\">I_{RHQ} = H_{SwinIR}(I_{LQ}) + I_{LQ}<\/code><\/pre>\n<p>\u5176\u4e2d <code class=\"katex-inline\">H_{SwinIR}(\\cdot)<\/code> \u8868\u793a SwinIR \u7684\u51fd\u6570\u3002<\/p>\n<h4>Loss function<\/h4>\n<p>\u5bf9\u4e8e\u56fe\u50cf\u8d85\u5206\u4efb\u52a1\uff0c\u6211\u4eec\u901a\u8fc7\u6700\u5c0f\u5316 <code class=\"katex-inline\">L_1<\/code> \u50cf\u7d20\u635f\u5931\u6765\u4f18\u5316 SwinIR \u7684\u53c2\u6570\uff1a<\/p>\n<pre><code class=\"language-katex\">\\mathcal{L} = \\| I_{RHQ} - I_{HQ} \\|_1<\/code><\/pre>\n<p>\u5176\u4e2d <code class=\"katex-inline\">I_{RHQ}<\/code> \u662f\u4ee5 <code class=\"katex-inline\">I_{LQ}<\/code> \u4e3a\u8f93\u5165\uff0c\u901a\u8fc7 SwinIR \u5f97\u5230\u7684\u7ed3\u679c\uff0c<code class=\"katex-inline\">I_{HQ}<\/code> \u662f\u5bf9\u5e94\u7684\u771f\u5b9e\u9ad8\u8d28\u91cf\u56fe\u50cf\u3002<\/p>\n<p>\u5bf9\u4e8e\u7ecf\u5178\u6216\u8f7b\u91cf\u7ea7\u56fe\u50cf\u8d85\u5206\u4efb\u52a1\uff0c\u6211\u4eec\u4ec5\u4f7f\u7528\u7b80\u5355\u7684 <code class=\"katex-inline\">L_1<\/code> \u50cf\u7d20\u635f\u5931\u4e0e\u4ee5\u5f80\u5de5\u4f5c\u4fdd\u6301\u4e00\u81f4\uff0c\u4ee5\u5c55\u793a\u6240\u63d0\u51fa\u7f51\u7edc\u7684\u6709\u6548\u6027\u3002\u5bf9\u4e8e\u771f\u5b9e\u56fe\u50cf\u8d85\u5206\u4efb\u52a1\uff0c\u6211\u4eec\u7ed3\u5408\u50cf\u7d20\u635f\u5931\u3001GAN \u635f\u5931\u4ee5\u53ca\u611f\u77e5\u635f\u5931 [81, 89, 80, 27, 39, 81] \u6765\u63d0\u5347\u89c6\u89c9\u8d28\u91cf\u3002<\/p>\n<p>\u5bf9\u4e8e\u56fe\u50cf\u53bb\u566a\u548c JPEG \u538b\u7f29\u4f2a\u5f71\u53bb\u9664\u4efb\u52a1\uff0c\u6211\u4eec\u4f7f\u7528 Charbonnier \u635f\u5931 [8]\uff0c\u5176\u5b9a\u4e49\u5982\u4e0b\uff1a<\/p>\n<pre><code class=\"language-katex\">\\mathcal{L} = \\sqrt{ \\| I_{RHQ} - I_{HQ} \\|^2 + \\epsilon^2 }<\/code><\/pre>\n<p>\u5176\u4e2d <code class=\"katex-inline\">\\epsilon<\/code> \u662f\u4e00\u4e2a\u5e38\u6570\uff0c\u7ecf\u9a8c\u4e0a\u8bbe\u4e3a <code class=\"katex-inline\">10^{-3}<\/code>\u3002<\/p>\n<h3>3.2. Residual Swin Transformer Block<\/h3>\n<p>\u5982\u56fe 2(a) \u6240\u793a\uff0c\u6b8b\u5dee Swin Transformer \u5757\uff08RSTB\uff09\u662f\u4e00\u4e2a\u5305\u542b Swin Transformer \u5c42\uff08STL\uff09\u548c\u5377\u79ef\u5c42\u7684\u6b8b\u5dee\u7ed3\u6784\u3002\u7ed9\u5b9a\u7b2c <code class=\"katex-inline\">i<\/code> \u4e2a RSTB \u7684\u8f93\u5165\u7279\u5f81 <code class=\"katex-inline\">F_{i,0}<\/code>\uff0c\u6211\u4eec\u9996\u5148\u901a\u8fc7 <code class=\"katex-inline\">L<\/code> \u4e2a Swin Transformer \u5c42\u63d0\u53d6\u4e2d\u95f4\u7279\u5f81 <code class=\"katex-inline\">F_{i,1}, F_{i,2}, \\dots, F_{i,L}<\/code>\uff0c\u8ba1\u7b97\u5982\u4e0b\uff1a<\/p>\n<pre><code class=\"language-katex\">F_{i,j} = H_{STL_{i,j}}(F_{i,j-1}), \\quad j = 1, 2, \\dots, L<\/code><\/pre>\n<p>\u5176\u4e2d <code class=\"katex-inline\">H_{STL_{i,j}}(\\cdot)<\/code> \u662f\u7b2c <code class=\"katex-inline\">i<\/code> \u4e2a RSTB \u4e2d\u7b2c <code class=\"katex-inline\">j<\/code> \u4e2a Swin Transformer \u5c42\u3002\u63a5\u7740\uff0c\u5728\u6b8b\u5dee\u8fde\u63a5\u524d\u6dfb\u52a0\u4e00\u4e2a\u5377\u79ef\u5c42\uff0cRSTB \u7684\u8f93\u51fa\u5b9a\u4e49\u4e3a\uff1a<\/p>\n<pre><code class=\"language-katex\">F_{i,\\text{out}} = H_{CONV_i}(F_{i,L}) + F_{i,0}<\/code><\/pre>\n<p>\u5176\u4e2d <code class=\"katex-inline\">H_{CONV_i}(\\cdot)<\/code> \u662f\u7b2c <code class=\"katex-inline\">i<\/code> \u4e2a RSTB \u4e2d\u7684\u5377\u79ef\u5c42\u3002\u8fd9\u4e00\u8bbe\u8ba1\u6709\u4e24\u4e2a\u597d\u5904\uff1a<\/p>\n<ol>\n<li>\u5c3d\u7ba1 Transformer \u53ef\u4ee5\u770b\u4f5c\u662f\u7a7a\u95f4\u53d8\u5316\u5377\u79ef\u7684\u4e00\u79cd\u7279\u4f8b [21, 75]\uff0c\u4f46\u5377\u79ef\u5c42\u5177\u6709\u7a7a\u95f4\u4e0d\u53d8\u6027\u7684\u6ee4\u6ce2\u5668\uff0c\u53ef\u4ee5\u589e\u5f3a SwinIR \u7684\u5e73\u79fb\u7b49\u53d8\u6027\uff1b<\/li>\n<li>\u6b8b\u5dee\u8fde\u63a5\u63d0\u4f9b\u4e86\u4e00\u79cd\u57fa\u4e8e\u6052\u7b49\u6620\u5c04\u7684\u8fde\u63a5\u65b9\u5f0f\uff0c\u5c06\u4e0d\u540c\u6a21\u5757\u7684\u4fe1\u606f\u4f20\u9012\u5230\u91cd\u5efa\u6a21\u5757\uff0c\u6709\u52a9\u4e8e\u878d\u5408\u4e0d\u540c\u5c42\u7ea7\u7684\u7279\u5f81\u3002<\/li>\n<\/ol>\n<h4>Swin Transformer layer<\/h4>\n<p>Swin Transformer \u5c42\uff08STL\uff09[56] \u57fa\u4e8e\u539f\u59cb Transformer \u5c42\u7684\u6807\u51c6\u591a\u5934\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08multi-head self-attention\uff09[76]\u3002\u5176\u4e3b\u8981\u533a\u522b\u5728\u4e8e\u5c40\u90e8\u6ce8\u610f\u529b\u548c\u6ed1\u52a8\u7a97\u53e3\u673a\u5236\u3002<\/p>\n<p>\u5982\u56fe 2(b) \u6240\u793a\uff0c\u7ed9\u5b9a\u8f93\u5165\u5c3a\u5bf8\u4e3a <code class=\"katex-inline\">H \\times W \\times C<\/code>\uff0cSwin Transformer \u9996\u5148\u5c06\u8f93\u5165\u5212\u5206\u4e3a\u4e0d\u91cd\u53e0\u7684 <code class=\"katex-inline\">M \\times M<\/code> \u5c40\u90e8\u7a97\u53e3\uff0c\u5f97\u5230\u5927\u5c0f\u4e3a <code class=\"katex-inline\">\\frac{HW}{M^2} \\times M^2 \\times C<\/code> \u7684\u7279\u5f81\u3002\u5176\u4e2d <code class=\"katex-inline\">\\frac{HW}{M^2}<\/code> \u662f\u7a97\u53e3\u603b\u6570\u3002\u968f\u540e\u5bf9\u6bcf\u4e2a\u7a97\u53e3\u5206\u522b\u6267\u884c\u6807\u51c6\u7684\u81ea\u6ce8\u610f\u529b\u64cd\u4f5c\uff08\u5373\u5c40\u90e8\u6ce8\u610f\u529b\uff09\u3002<\/p>\n<p>\u5bf9\u4e8e\u4e00\u4e2a\u5c40\u90e8\u7a97\u53e3\u7279\u5f81 <code class=\"katex-inline\">X \\in \\mathbb{R}^{M^2 \\times C}<\/code>\uff0c\u67e5\u8be2\uff08query\uff09\u3001\u952e\uff08key\uff09\u548c\u503c\uff08value\uff09\u77e9\u9635 <code class=\"katex-inline\">Q<\/code>\u3001<code class=\"katex-inline\">K<\/code> \u548c <code class=\"katex-inline\">V<\/code> \u7684\u8ba1\u7b97\u5982\u4e0b\uff1a<\/p>\n<pre><code class=\"language-katex\">Q = XP_Q, \\quad K = XP_K, \\quad V = XP_V<\/code><\/pre>\n<p>\u5176\u4e2d <code class=\"katex-inline\">P_Q<\/code>\u3001<code class=\"katex-inline\">P_K<\/code> \u548c <code class=\"katex-inline\">P_V<\/code> \u662f\u6295\u5f71\u77e9\u9635\uff0c\u5728\u4e0d\u540c\u7a97\u53e3\u4e4b\u95f4\u5171\u4eab\u3002\u901a\u5e38\u60c5\u51b5\u4e0b\uff0c<code class=\"katex-inline\">Q, K, V \\in \\mathbb{R}^{M^2 \\times d}<\/code>\u3002\u6ce8\u610f\u529b\u77e9\u9635\u901a\u8fc7\u5c40\u90e8\u7a97\u53e3\u5185\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u8ba1\u7b97\u5982\u4e0b\uff1a<\/p>\n<pre><code class=\"language-katex\">\\text{Attention}(Q, K, V) = \\text{SoftMax}\\left(\\frac{QK^T}{\\sqrt{d}} + B\\right)V<\/code><\/pre>\n<p>\u5176\u4e2d <code class=\"katex-inline\">B<\/code> \u662f\u53ef\u5b66\u4e60\u7684\u76f8\u5bf9\u4f4d\u7f6e\u7f16\u7801\u3002\u5728\u5b9e\u9645\u64cd\u4f5c\u4e2d\uff0c\u9075\u5faa [76] \u7684\u505a\u6cd5\uff0c\u6211\u4eec\u5e76\u884c\u6267\u884c <code class=\"katex-inline\">h<\/code> \u6b21\u6ce8\u610f\u529b\u8ba1\u7b97\uff0c\u5e76\u5c06\u7ed3\u679c\u62fc\u63a5\u7528\u4e8e\u591a\u5934\u81ea\u6ce8\u610f\u529b\uff08MSA\uff09\u3002<\/p>\n<p>\u63a5\u4e0b\u6765\u662f\u4e00\u4e2a\u591a\u5c42\u611f\u77e5\u673a\uff08MLP\uff09\uff0c\u5b83\u5305\u542b\u4e24\u4e2a\u5e26\u6709 GELU \u975e\u7ebf\u6027\u6fc0\u6d3b\u7684\u5168\u8fde\u63a5\u5c42\uff0c\u7528\u4e8e\u8fdb\u4e00\u6b65\u7684\u7279\u5f81\u53d8\u6362\u3002\u5728 MSA \u548c MLP \u4e4b\u524d\u90fd\u6dfb\u52a0\u4e86 LayerNorm\uff08LN\uff09\u5c42\uff0c\u4e14\u4e24\u4e2a\u6a21\u5757\u90fd\u4f7f\u7528\u4e86\u6b8b\u5dee\u8fde\u63a5\uff0c\u6574\u4e2a\u8fc7\u7a0b\u5982\u4e0b\uff1a<\/p>\n<pre><code class=\"language-katex\">X = \\text{MSA}(\\text{LN}(X)) + X,<\/code><\/pre>\n<pre><code class=\"language-katex\">X = \\text{MLP}(\\text{LN}(X)) + X<\/code><\/pre>\n<p>\u7136\u800c\uff0c\u5f53\u6bcf\u5c42\u4f7f\u7528\u56fa\u5b9a\u5212\u5206\u65b9\u5f0f\u65f6\uff0c\u4e0d\u540c\u7a97\u53e3\u4e4b\u95f4\u65e0\u6cd5\u5efa\u7acb\u8054\u7cfb\u3002\u56e0\u6b64\uff0c\u4ea4\u66ff\u4f7f\u7528\u89c4\u5219\u5212\u5206\u548c\u6ed1\u52a8\u7a97\u53e3\u5212\u5206\u6765\u5b9e\u73b0\u8de8\u7a97\u53e3\u8fde\u63a5 [56]\u3002\u6ed1\u52a8\u7a97\u53e3\u5212\u5206\u7684\u542b\u4e49\u662f\uff1a\u5728\u5212\u5206\u524d\u5c06\u7279\u5f81\u56fe\u6cbf\u4e24\u4e2a\u65b9\u5411\u5206\u522b\u5e73\u79fb <code class=\"katex-inline\">\\left(\\left\\lfloor \\frac{M}{2} \\right\\rfloor, \\left\\lfloor \\frac{M}{2} \\right\\rfloor\\right)<\/code> \u4e2a\u50cf\u7d20\u3002<\/p>\n<h2>4. Experiment<\/h2>\n<h3>4.1. Experimental Setup<\/h3>\n<p>\u5bf9\u4e8e\u7ecf\u5178\u56fe\u50cf\u8d85\u5206\u3001\u771f\u5b9e\u56fe\u50cf\u8d85\u5206\u3001\u56fe\u50cf\u53bb\u566a\u548c JPEG \u538b\u7f29\u4f2a\u5f71\u53bb\u9664\u7b49\u4efb\u52a1\uff0cRSTB \u6570\u91cf\u3001STL \u6570\u91cf\u3001\u7a97\u53e3\u5927\u5c0f\u3001\u901a\u9053\u6570\u548c\u6ce8\u610f\u529b\u5934\u6570\u901a\u5e38\u8bbe\u7f6e\u4e3a 6\u30016\u30018\u3001180 \u548c 6\u3002\u5176\u4e2d\u4e00\u4e2a\u4f8b\u5916\u662f\uff1a\u5728 JPEG \u538b\u7f29\u4f2a\u5f71\u53bb\u9664\u4efb\u52a1\u4e2d\uff0c\u7a97\u53e3\u5927\u5c0f\u8bbe\u7f6e\u4e3a 7\uff0c\u56e0\u4e3a\u6211\u4eec\u5728\u4f7f\u7528 8 \u65f6\u6027\u80fd\u663e\u8457\u4e0b\u964d\uff0c\u53ef\u80fd\u662f\u56e0\u4e3a JPEG \u7f16\u7801\u91c7\u7528\u7684\u662f <code class=\"katex-inline\">8 \\times 8<\/code> \u7684\u56fe\u50cf\u5206\u5757\u65b9\u5f0f\u3002<\/p>\n<p>\u5bf9\u4e8e\u8f7b\u91cf\u7ea7\u56fe\u50cf\u8d85\u5206\u4efb\u52a1\uff0c\u6211\u4eec\u5c06 RSTB \u6570\u91cf\u548c\u901a\u9053\u6570\u5206\u522b\u51cf\u5c11\u4e3a 4 \u548c 60\u3002\u53c2\u8003 [95, 63]\uff0c\u5728\u6d4b\u8bd5\u8fc7\u7a0b\u4e2d\u4f7f\u7528\u81ea\u96c6\u6210\u7b56\u7565\uff08self-ensemble strategy\uff09[51] \u65f6\uff0c\u6211\u4eec\u4f7f\u7528\u7b26\u53f7 \u201c+\u201d \u8868\u793a\u6a21\u578b\uff0c\u4f8b\u5982 SwinIR+\u3002\u7531\u4e8e\u7bc7\u5e45\u9650\u5236\uff0c\u8bad\u7ec3\u4e0e\u8bc4\u4f30\u7ec6\u8282\u5728\u8865\u5145\u6750\u6599\u4e2d\u63d0\u4f9b\u3002<\/p>\n<h3>4.2. Ablation Study and Discussion<\/h3>\n<p>\u4e3a\u4e86\u8fdb\u884c\u6d88\u878d\u5b9e\u9a8c\uff0c\u6211\u4eec\u5728 DIV2K \u6570\u636e\u96c6\u4e0a\u8bad\u7ec3 SwinIR\uff08\u00d72 \u653e\u5927\uff09\u5e76\u5728 Manga109 \u4e0a\u8fdb\u884c\u6d4b\u8bd5 [1, 60]\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/pic.cirno.fun\/sssn-blog-pics\/image-20250728012750463.png\" alt=\"image-20250728012750463\" \/><\/p>\n<h4>Impact of channel number, RSTB number and STL number<\/h4>\n<p>\u6211\u4eec\u5728 RSTB \u4e2d\u7814\u7a76\u4e86\u901a\u9053\u6570\u3001RSTB \u6570\u91cf\u548c Swin Transformer \u5c42\u6570\u5bf9\u6a21\u578b\u6027\u80fd\u7684\u5f71\u54cd\uff0c\u7ed3\u679c\u5982\u56fe 3(a)\u30013(b)\u30013(c)\u3002\u5b9e\u9a8c\u8868\u660e\uff0cPSNR \u4e0e\u8fd9\u4e09\u4e2a\u8d85\u53c2\u6570\u5448\u6b63\u76f8\u5173\u3002<\/p>\n<ul>\n<li>\u901a\u9053\u6570\uff1a\u968f\u7740\u901a\u9053\u6570\u589e\u52a0\uff0c\u6027\u80fd\u6301\u7eed\u63d0\u5347\uff0c\u4f46\u53c2\u6570\u91cf\u5448\u5e73\u65b9\u589e\u957f\u3002\u4e3a\u4e86\u5728\u6027\u80fd\u4e0e\u6a21\u578b\u89c4\u6a21\u4e4b\u95f4\u53d6\u5f97\u5e73\u8861\uff0c\u6211\u4eec\u5728\u5176\u4f59\u5b9e\u9a8c\u4e2d\u5c06\u901a\u9053\u6570\u8bbe\u4e3a 180\u3002<\/li>\n<li>RSTB \u6570\u91cf\u4e0e\u5c42\u6570\uff1a\u6027\u80fd\u63d0\u5347\u9010\u6e10\u8d8b\u4e8e\u9971\u548c\uff0c\u6700\u7ec8\u9009\u62e9\u4e8c\u8005\u5747\u4e3a 6\uff0c\u4ee5\u6784\u5efa\u76f8\u5bf9\u8f83\u5c0f\u7684\u6a21\u578b\u3002<\/li>\n<\/ul>\n<h4>Impact of patch size and training image number; model convergence comparison<\/h4>\n<p>\u6211\u4eec\u5c06 SwinIR \u4e0e\u5178\u578b CNN \u6a21\u578b RCAN \u8fdb\u884c\u5bf9\u6bd4\uff0c\u7814\u7a76 Transformer \u4e0e CNN \u6a21\u578b\u7684\u5dee\u5f02\u3002<\/p>\n<ul>\n<li>\u4ece\u56fe 3(d) \u53ef\u770b\u51fa\uff0cSwinIR \u5728\u4e0d\u540c patch size \u8bbe\u7f6e\u4e0b\u8868\u73b0\u5747\u4f18\u4e8e RCAN\uff0cpatch \u8d8a\u5927\uff0cPSNR \u63d0\u5347\u8d8a\u660e\u663e\u3002<\/li>\n<li>\u56fe 3(e) \u663e\u793a\u8bad\u7ec3\u56fe\u50cf\u6570\u91cf\u7684\u5f71\u54cd\uff0c\u4f7f\u7528\u989d\u5916\u7684 Flickr2K \u56fe\u50cf\u8fdb\u884c\u8bad\u7ec3\u3002\u7ed3\u679c\u8868\u660e\uff1a\n<ul>\n<li>SwinIR \u6027\u80fd\u968f\u7740\u8bad\u7ec3\u56fe\u50cf\u6570\u91cf\u589e\u52a0\u800c\u63d0\u5347\uff1b<\/li>\n<li>\u4e0e IPT \u4e2d Transformer \u6a21\u578b\u5bf9\u5927\u91cf\u6570\u636e\u7684\u4f9d\u8d56\u4e0d\u540c\uff0cSwinIR \u5728\u8bad\u7ec3\u6570\u636e\u8f83\u5c11\uff08\u5982 25%\uff0c200 \u5f20\u56fe\u50cf\uff09\u65f6\u4e5f\u4f18\u4e8e CNN \u6a21\u578b\uff1b<\/li>\n<\/ul>\n<\/li>\n<li>\u56fe 3(f) \u5c55\u793a\u4e86 SwinIR \u4e0e RCAN \u7684\u8bad\u7ec3\u8fc7\u7a0b\u4e2d PSNR \u53d8\u5316\u66f2\u7ebf\uff0c\u663e\u793a SwinIR \u6536\u655b\u66f4\u5feb\uff0c\u6027\u80fd\u66f4\u597d\uff0c\u6253\u7834\u4e86 Transformer \u6a21\u578b\u6536\u655b\u6162\u7684\u5e38\u89c4\u8ba4\u77e5\u3002<\/li>\n<\/ul>\n<h4>Impact of residual connection and convolution layer in RSTB<\/h4>\n<p><img decoding=\"async\" src=\"https:\/\/pic.cirno.fun\/sssn-blog-pics\/image-20250728012820632.png\" alt=\"image-20250728012820632\" style=\"zoom:50%;\" \/><\/p>\n<p>\u8868 1 \u5c55\u793a\u4e86 RSTB \u4e2d\u56db\u79cd\u4e0d\u540c\u7684\u6b8b\u5dee\u8fde\u63a5\u53d8\u4f53\uff1a<\/p>\n<ul>\n<li>\u65e0\u6b8b\u5dee\u8fde\u63a5\uff1b<\/li>\n<li>\u4f7f\u7528 <code class=\"katex-inline\">1 \\times 1<\/code> \u5377\u79ef\uff1b<\/li>\n<li>\u4f7f\u7528 <code class=\"katex-inline\">3 \\times 3<\/code> \u5377\u79ef\uff1b<\/li>\n<li>\u4f7f\u7528\u4e09\u4e2a <code class=\"katex-inline\">3 \\times 3<\/code> \u5377\u79ef\uff08\u4e2d\u95f4\u901a\u9053\u4e3a\u603b\u901a\u9053\u6570\u7684\u56db\u5206\u4e4b\u4e00\uff09\u3002<\/li>\n<\/ul>\n<p>\u7ed3\u8bba\u5982\u4e0b\uff1a<\/p>\n<ol>\n<li>\u6b8b\u5dee\u8fde\u63a5\u5bf9\u6027\u80fd\u6709\u76ca\uff0c\u53ef\u63d0\u5347 PSNR \u7ea6 0.16dB\uff1b<\/li>\n<li><code class=\"katex-inline\">1 \\times 1<\/code> \u5377\u79ef\u63d0\u5347\u6709\u9650\uff0c\u53ef\u80fd\u662f\u56e0\u4e3a\u65e0\u6cd5\u50cf <code class=\"katex-inline\">3 \\times 3<\/code> \u5377\u79ef\u90a3\u6837\u63d0\u53d6\u90bb\u57df\u4fe1\u606f\uff1b<\/li>\n<li>\u4f7f\u7528\u4e09\u4e2a <code class=\"katex-inline\">3 \\times 3<\/code> \u5377\u79ef\u53ef\u51cf\u5c11\u53c2\u6570\u6570\u76ee\uff0c\u4f46\u6027\u80fd\u7565\u6709\u4e0b\u964d\u3002<\/li>\n<\/ol>\n<h3>4.3. Results on Image SR<\/h3>\n<h4>Classical image SR<\/h4>\n<p><img decoding=\"async\" src=\"https:\/\/pic.cirno.fun\/sssn-blog-pics\/image-20250728012855145.png\" alt=\"image-20250728012855145\" style=\"zoom:50%;\" \/><\/p>\n<p>\u8868 2 \u663e\u793a\u4e86 SwinIR\uff08\u4e2d\u7b49\u89c4\u6a21\uff09\u4e0e\u5f53\u524d\u6700\u4f18\u65b9\u6cd5\u7684\u5b9a\u91cf\u5bf9\u6bd4\uff0c\u5305\u62ec DBPN [31]\u3001RCAN [95]\u3001RRDB [81]\u3001SAN [15]\u3001IGNN [100]\u3001HAN [63]\u3001NLSA [61] \u548c IPT [9]\u3002<\/p>\n<ul>\n<li>\u5728 DIV2K \u4e0a\u8bad\u7ec3\u540e\uff0cSwinIR \u5728\u51e0\u4e4e\u6240\u6709\u57fa\u51c6\u6570\u636e\u96c6\u548c\u5c3a\u5ea6\u4e0a\u90fd\u53d6\u5f97\u6700\u4f73\u6027\u80fd\uff1b<\/li>\n<li>\u5728 Manga109 \u4e0a\uff0c\u653e\u5927\u56e0\u5b50\u4e3a 4 \u7684 PSNR \u6700\u5927\u63d0\u5347\u8fbe 0.26dB\uff1b<\/li>\n<li>\u5c3d\u7ba1 RCAN \u548c HAN \u5f15\u5165\u4e86\u901a\u9053\u4e0e\u7a7a\u95f4\u6ce8\u610f\u529b\u673a\u5236\uff0cIGNN \u4f7f\u7528\u4e86\u81ea\u9002\u5e94 patch \u7279\u5f81\u805a\u5408\uff0cNLSA \u91c7\u7528\u975e\u5c40\u90e8\u6ce8\u610f\u529b\uff0c\u4f46\u5b83\u4eec\u90fd\u4e0d\u5982\u57fa\u4e8e Transformer \u7684 SwinIR\uff1b<\/li>\n<li>\u4f7f\u7528\u66f4\u5927\u6570\u636e\u96c6\uff08DIV2K+Flickr2K\uff09\u8bad\u7ec3\u540e\uff0cSwinIR \u6027\u80fd\u8fdb\u4e00\u6b65\u63d0\u5347\uff0c\u6700\u9ad8\u63d0\u5347 0.47dB\uff0c\u8d85\u8fc7\u4e86\u4f7f\u7528 ImageNet\uff08130 \u4e07\u56fe\u50cf\uff09\u8bad\u7ec3\u5e76\u62e5\u6709 115.5M \u53c2\u6570\u7684 IPT \u6a21\u578b\uff1b<\/li>\n<li>\u76f8\u6bd4\u4e4b\u4e0b\uff0cSwinIR \u4ec5\u6709 11.8M \u53c2\u6570\uff0c\u6bd4 CNN \u6a21\u578b\uff0815.4~44.3M\uff09\u66f4\u5c0f\uff1b<\/li>\n<li>\u5728\u6d4b\u8bd5\u65f6\u95f4\u4e0a\uff0cRCAN\u3001IPT \u548c SwinIR \u5206\u522b\u4e3a 0.2s\u30014.5s \u548c 1.1s\uff08\u5728 <code class=\"katex-inline\">1024 \\times 1024<\/code> \u56fe\u50cf\u4e0a\uff09\uff0c\u5982\u56fe 4 \u6240\u793a\uff0cSwinIR \u80fd\u6062\u590d\u9ad8\u9891\u7ec6\u8282\u5e76\u51cf\u5c11\u6a21\u7cca\u4f2a\u5f71\uff0c\u751f\u6210\u6e05\u6670\u81ea\u7136\u7684\u56fe\u50cf\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"https:\/\/pic.cirno.fun\/sssn-blog-pics\/image-20250728013016600.png\" alt=\"image-20250728013016600\" \/><\/p>\n<h4>Lightweight image SR<\/h4>\n<p><img decoding=\"async\" src=\"https:\/\/pic.cirno.fun\/sssn-blog-pics\/image-20250728013033030.png\" alt=\"image-20250728013033030\" style=\"zoom:50%;\" \/><\/p>\n<p>\u6211\u4eec\u8fd8\u5c06 SwinIR\uff08\u5c0f\u89c4\u6a21\uff09\u4e0e\u5f53\u524d\u8f7b\u91cf\u7ea7 SR \u65b9\u6cd5\u8fdb\u884c\u5bf9\u6bd4\uff1aCARN [2]\u3001FALSR-A [12]\u3001IMDN [35]\u3001LAPAR-A [44] \u548c LatticeNet [57]\u3002<\/p>\n<ul>\n<li>\u9664\u4e86 PSNR \u548c SSIM\uff0c\u6211\u4eec\u8fd8\u62a5\u544a\u4e86\u53c2\u6570\u6570\u91cf\u548c\u4e58\u52a0\u8fd0\u7b97\u6b21\u6570\uff08\u5728 <code class=\"katex-inline\">1280 \\times 720<\/code> HQ \u56fe\u50cf\u4e0a\u8bc4\u4f30\uff09\uff1b<\/li>\n<li>\u5982\u8868 3 \u6240\u793a\uff0cSwinIR \u5728\u591a\u4e2a\u57fa\u51c6\u6570\u636e\u96c6\u4e0a\u6700\u591a\u63d0\u9ad8 0.53dB\uff0c\u4e14\u53c2\u6570\u91cf\u4e0e\u8ba1\u7b97\u91cf\u76f8\u8fd1\uff1b<\/li>\n<li>\u8bf4\u660e SwinIR \u67b6\u6784\u5728\u56fe\u50cf\u590d\u539f\u4efb\u52a1\u4e2d\u5177\u6709\u6781\u9ad8\u6548\u7387\u3002<\/li>\n<\/ul>\n<h4>Real-world image SR<\/h4>\n<p><img decoding=\"async\" src=\"https:\/\/pic.cirno.fun\/sssn-blog-pics\/image-20250728013115399.png\" alt=\"image-20250728013115399\" style=\"zoom:50%;\" \/><\/p>\n<p>\u56fe\u50cf\u8d85\u5206\u7684\u6700\u7ec8\u76ee\u6807\u662f\u5b9e\u9645\u5e94\u7528\u3002\u8fd1\u671f\uff0cZhang \u7b49\u4eba [89] \u63d0\u51fa\u5b9e\u7528\u7684\u9000\u5316\u6a21\u578b BSRGAN\uff0c\u5e76\u5728\u771f\u5b9e\u56fe\u50cf SR \u4efb\u52a1\u4e2d\u8868\u73b0\u4f18\u5f02\u3002<\/p>\n<ul>\n<li>\u4e3a\u8bc4\u4f30 SwinIR \u5728\u771f\u5b9e\u56fe\u50cf SR \u7684\u6027\u80fd\uff0c\u6211\u4eec\u4f7f\u7528\u4e0e BSRGAN \u76f8\u540c\u7684\u9000\u5316\u6a21\u578b\u5bf9 SwinIR \u8fdb\u884c\u518d\u8bad\u7ec3\uff1b<\/li>\n<li>\u7531\u4e8e\u6ca1\u6709\u771f\u5b9e\u9ad8\u8d28\u91cf\u56fe\u50cf\uff0c\u6211\u4eec\u4ec5\u4e0e\u4ee3\u8868\u6027\u6a21\u578b\uff08\u5982 ESRGAN [81]\u3001RealSR [37]\u3001BSRGAN [89]\u3001Real-ESRGAN [80]\uff09\u8fdb\u884c\u89c6\u89c9\u5bf9\u6bd4\uff1b<\/li>\n<li>\u5982\u56fe 5 \u6240\u793a\uff0cSwinIR \u751f\u6210\u7684\u56fe\u50cf\u5177\u6709\u6e05\u6670\u9510\u5229\u7684\u8fb9\u7f18\u548c\u826f\u597d\u7684\u611f\u5b98\u8d28\u91cf\uff0c\u800c\u5176\u4ed6\u65b9\u6cd5\u53ef\u80fd\u5b58\u5728\u4f2a\u5f71\uff1b<\/li>\n<li>\u4e3a\u8fdb\u4e00\u6b65\u53d1\u6398 SwinIR \u5728\u771f\u5b9e\u573a\u666f\u4e2d\u7684\u6f5c\u529b\uff0c\u6211\u4eec\u8bad\u7ec3\u4e86\u4e00\u4e2a\u66f4\u5927\u7684\u6a21\u578b\uff0c\u5e76\u5728\u66f4\u5927\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n<p>\u5b9e\u9a8c\u8868\u660e\uff0c\u8be5\u5927\u6a21\u578b\u80fd\u591f\u5904\u7406\u66f4\u590d\u6742\u7684\u9000\u5316\u95ee\u9898\uff0c\u5728\u771f\u5b9e\u56fe\u50cf\u4e0a\u8868\u73b0\u66f4\u4f73\u3002\u8be6\u7ec6\u5185\u5bb9\u8bf7\u53c2\u89c1\u9879\u76ee\u4e3b\u9875\uff1a <a href=\"https:\/\/github.com\/JingyunLiang\/SwinIR\">https:\/\/github.com\/JingyunLiang\/SwinIR<\/a><\/p>\n<h3>4.4. Results on JPEG Compression Artifact Reduction<\/h3>\n<p><img decoding=\"async\" src=\"https:\/\/pic.cirno.fun\/sssn-blog-pics\/image-20250728013157289.png\" alt=\"image-20250728013157289\" style=\"zoom:50%;\" \/><\/p>\n<p>\u8868 4 \u5c55\u793a\u4e86 SwinIR \u4e0e\u5f53\u524d\u6700\u5148\u8fdb\u7684 JPEG \u538b\u7f29\u4f2a\u5f71\u53bb\u9664\u65b9\u6cd5\u7684\u5bf9\u6bd4\uff0c\u5305\u62ec\uff1aARCNN [17]\u3001DnCNN-3 [90]\u3001QGAC [20]\u3001RNAN [96]\u3001RDN [98] \u548c DRUNet [88]\u3002\u8fd9\u4e9b\u65b9\u6cd5\u5168\u90e8\u57fa\u4e8e CNN\u3002<\/p>\n<p>\u6309\u7167 [98, 88] \u7684\u8bbe\u7f6e\uff0c\u6211\u4eec\u5728\u4e24\u4e2a\u57fa\u51c6\u6570\u636e\u96c6\uff08Classic5 [22] \u548c LIVE1 [67]\uff09\u4e0a\u6d4b\u8bd5\u4e0d\u540c\u65b9\u6cd5\uff0cJPEG \u56fe\u50cf\u8d28\u91cf\u56e0\u5b50\u5206\u522b\u8bbe\u4e3a 10\u300120\u300130 \u548c 40\u3002<\/p>\n<p>\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u6240\u63d0\u51fa\u7684 SwinIR \u5728\u4e0d\u540c\u8d28\u91cf\u56e0\u5b50\u4e0b\uff0c\u5728\u4e24\u4e2a\u6d4b\u8bd5\u96c6\u4e0a\u7684\u5e73\u5747 PSNR \u81f3\u5c11\u63d0\u5347\u4e86 0.11dB \u548c 0.07dB\u3002\u6b64\u5916\uff0cSwinIR \u4ec5\u6709 11.5M \u53c2\u6570\uff0c\u800c\u6b64\u524d\u6700\u4f73\u6a21\u578b DRUNet \u62e5\u6709 32.7M \u53c2\u6570\uff0c\u663e\u8457\u66f4\u8f7b\u91cf\u3002<\/p>\n<h3>4.5. Results on Image Denoising<\/h3>\n<p><img decoding=\"async\" src=\"https:\/\/pic.cirno.fun\/sssn-blog-pics\/image-20250728013253719.png\" alt=\"image-20250728013253719\" style=\"zoom:50%;\" \/><\/p>\n<p>\u6211\u4eec\u5728\u8868 5 \u548c\u8868 6 \u4e2d\u5206\u522b\u5c55\u793a\u4e86\u7070\u5ea6\u56fe\u50cf\u548c\u5f69\u8272\u56fe\u50cf\u7684\u53bb\u566a\u7ed3\u679c\u3002<\/p>\n<p>\u5bf9\u6bd4\u65b9\u6cd5\u5305\u62ec\u4f20\u7edf\u65b9\u6cd5\uff08BM3D [14] \u548c WNNM [29]\uff09\u3001CNN \u6a21\u578b\uff08\u5982 DnCNN [90]\u3001IRCNN [91]\u3001FFDNet [92]\u3001N3Net [65]\u3001NLRN [52]\u3001FOCNet [38]\u3001RNAN [96]\u3001MWCNN [54] \u548c DRUNet [88]\uff09\u3002<\/p>\n<p>\u6839\u636e [90, 88]\uff0c\u6211\u4eec\u5728\u566a\u58f0\u6c34\u5e73\u4e3a 15\u300125 \u548c 50 \u7684\u60c5\u51b5\u4e0b\u8fdb\u884c\u5bf9\u6bd4\u3002<\/p>\n<p>\u5b9e\u9a8c\u8868\u660e\uff0cSwinIR \u5728\u6240\u6709\u5bf9\u6bd4\u65b9\u6cd5\u4e2d\u8868\u73b0\u6700\u597d\uff0c\u5c24\u5176\u662f\u5728\u5305\u542b 100 \u5f20\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u7684 Urban100 \u6570\u636e\u96c6\u4e0a\uff0cPSNR \u63d0\u5347\u9ad8\u8fbe 0.3dB\uff0c\u8d85\u8fc7\u5f53\u524d\u6700\u4f18\u6a21\u578b DRUNet\u3002<\/p>\n<p>\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cSwinIR \u4ec5\u6709 12.0M \u53c2\u6570\uff0c\u800c DRUNet \u6709 32.7M \u53c2\u6570\uff0c\u8fd9\u8868\u660e SwinIR \u5728\u5b66\u4e60\u7528\u4e8e\u56fe\u50cf\u590d\u539f\u7684\u7279\u5f81\u8868\u793a\u65b9\u9762\u5177\u6709\u9ad8\u5ea6\u6548\u7387\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/pic.cirno.fun\/sssn-blog-pics\/image-20250728013309227.png\" alt=\"image-20250728013309227\" style=\"zoom:50%;\" \/><\/p>\n<p>\u7070\u5ea6\u548c\u5f69\u8272\u56fe\u50cf\u7684\u53ef\u89c6\u5316\u5bf9\u6bd4\u7ed3\u679c\u5982\u56fe 6 \u548c\u56fe 7 \u6240\u793a\u3002\u5b9e\u9a8c\u8868\u660e\uff0cSwinIR \u80fd\u591f\u6709\u6548\u53bb\u9664\u4e25\u91cd\u566a\u58f0\u6270\u52a8\uff0c\u4fdd\u7559\u56fe\u50cf\u4e2d\u7684\u9ad8\u9891\u7ec6\u8282\uff0c\u4ece\u800c\u83b7\u5f97\u66f4\u6e05\u6670\u7684\u8fb9\u7f18\u548c\u81ea\u7136\u7684\u7eb9\u7406\u3002\u800c\u5176\u4ed6\u65b9\u6cd5\u8981\u4e48\u8fc7\u4e8e\u5e73\u6ed1\uff0c\u8981\u4e48\u9510\u5316\u8fc7\u5ea6\uff0c\u65e0\u6cd5\u6062\u590d\u4e30\u5bcc\u7684\u7eb9\u7406\u4fe1\u606f\u3002<\/p>\n<h2>5. Conclusion<\/h2>\n<p>\u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e Swin Transformer \u7684\u56fe\u50cf\u590d\u539f\u6a21\u578b\u2014\u2014SwinIR\u3002\u8be5\u6a21\u578b\u7531\u4e09\u4e2a\u90e8\u5206\u7ec4\u6210\uff1a\u6d45\u5c42\u7279\u5f81\u63d0\u53d6\u3001\u6df1\u5c42\u7279\u5f81\u63d0\u53d6\u548c\u9ad8\u8d28\u91cf\u56fe\u50cf\u91cd\u5efa\u6a21\u5757\u3002<\/p>\n<p>\u7279\u522b\u5730\uff0c\u6211\u4eec\u4e3a\u6df1\u5c42\u7279\u5f81\u63d0\u53d6\u4f7f\u7528\u4e86\u4e00\u7ec4\u6b8b\u5dee Swin Transformer \u5757\uff08RSTB\uff09\uff0c\u6bcf\u4e2a RSTB \u7531 Swin Transformer \u5c42\u3001\u5377\u79ef\u5c42\u548c\u6b8b\u5dee\u8fde\u63a5\u7ec4\u6210\u3002<\/p>\n<p>\u5927\u91cf\u5b9e\u9a8c\u8bc1\u660e\uff0cSwinIR \u5728\u4e09\u7c7b\u5177\u6709\u4ee3\u8868\u6027\u7684\u56fe\u50cf\u590d\u539f\u4efb\u52a1\u4e2d\uff0c\u5728\u516d\u79cd\u4e0d\u540c\u8bbe\u7f6e\u4e0b\u5747\u8fbe\u5230\u4e86\u5f53\u524d\u6700\u5148\u8fdb\u7684\u6027\u80fd\u3002\u8fd9\u4e9b\u8bbe\u7f6e\u5305\u62ec\uff1a\u7ecf\u5178\u56fe\u50cf\u8d85\u5206\u8fa8\uff08SR\uff09\u3001\u8f7b\u91cf\u7ea7\u56fe\u50cf SR\u3001\u771f\u5b9e\u56fe\u50cf SR\u3001\u7070\u5ea6\u56fe\u50cf\u53bb\u566a\u3001\u5f69\u8272\u56fe\u50cf\u53bb\u566a\u548c JPEG \u538b\u7f29\u4f2a\u5f71\u53bb\u9664\u3002\u8fd9\u5145\u5206\u5c55\u793a\u4e86\u6240\u63d0\u51fa\u7684 SwinIR \u7684\u6709\u6548\u6027\u548c\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n<p>\u672a\u6765\uff0c\u6211\u4eec\u5c06\u6269\u5c55\u8be5\u6a21\u578b\u81f3\u5176\u4ed6\u56fe\u50cf\u590d\u539f\u4efb\u52a1\uff0c\u4f8b\u5982\u56fe\u50cf\u53bb\u6a21\u7cca\u548c\u53bb\u96e8\u7b49\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>SwinIR: Image Restoration Using Swin Tra [&#8230;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[32],"tags":[],"class_list":["post-859","post","type-post","status-publish","format-standard","hentry","category-32"],"_links":{"self":[{"href":"https:\/\/blog.sssn.tech\/index.php?rest_route=\/wp\/v2\/posts\/859","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.sssn.tech\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.sssn.tech\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.sssn.tech\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.sssn.tech\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=859"}],"version-history":[{"count":3,"href":"https:\/\/blog.sssn.tech\/index.php?rest_route=\/wp\/v2\/posts\/859\/revisions"}],"predecessor-version":[{"id":862,"href":"https:\/\/blog.sssn.tech\/index.php?rest_route=\/wp\/v2\/posts\/859\/revisions\/862"}],"wp:attachment":[{"href":"https:\/\/blog.sssn.tech\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=859"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.sssn.tech\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=859"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.sssn.tech\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=859"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}