{"id":2263,"date":"2026-07-15T23:11:39","date_gmt":"2026-07-15T15:11:39","guid":{"rendered":"https:\/\/ferina.top\/?p=2263"},"modified":"2026-07-15T23:11:56","modified_gmt":"2026-07-15T15:11:56","slug":"dlss","status":"publish","type":"post","link":"https:\/\/ferina.top\/index.php\/2026\/07\/15\/dlss\/","title":{"rendered":"DLSS"},"content":{"rendered":"\n

1. DLSS \u603b\u89c8<\/h2>\n\n\n\n

\u7531\u4e8eNVIDIA\u7684\u5ba3\u4f20\u95ee\u9898\uff0cDLSS\u5b9e\u9645\u5e76\u4e0d\u4ec5\u4ec5\u662f\u72ed\u4e49SS\uff08super sampling\uff09\u3002\u5b83\u5b9e\u9645\u5305\u542b\u4e86\u4ee5\u4e0b\u56db\u4e2a\u65b9\u9762\uff1a<\/p>\n\n\n\n

\u6280\u672f<\/th>\u4e3b\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/th>\u8f93\u5165<\/th>\u6838\u5fc3\u8f93\u51fa<\/th><\/tr><\/thead>
DLSS Super Resolution<\/td>\u5206\u8fa8\u7387\u4e0d\u8db3\u3001\u6297\u952f\u9f7f<\/td>\u4f4e\u5206\u8fa8\u7387\u753b\u9762\u3001\u8fd0\u52a8\u77e2\u91cf\u3001\u5386\u53f2\u5e27\u7b49<\/td>\u66f4\u9ad8\u5206\u8fa8\u7387\u753b\u9762<\/td><\/tr>
DLSS Ray Reconstruction<\/td>\u5149\u8ffd\u6837\u672c\u592a\u5c11\u3001\u566a\u58f0\u4e25\u91cd<\/td>\u7a00\u758f\u5149\u7ebf\u6837\u672c\u3001G-buffer\u3001\u5386\u53f2\u4fe1\u606f<\/td>\u53bb\u566a\u5e76\u91cd\u5efa\u53cd\u5c04\u3001\u9634\u5f71\u3001GI<\/td><\/tr>
DLSS Frame Generation<\/td>\u663e\u793a\u5e27\u7387\u4e0d\u8db3<\/td>\u524d\u540e\u5e27\u3001\u8fd0\u52a8\/\u5149\u6d41\u4fe1\u606f<\/td>\u989d\u5916\u7684\u4e2d\u95f4\u5e27<\/td><\/tr>
DLSS 5<\/strong><\/td>\u4f20\u7edf\u5b9e\u65f6\u7740\u8272\u96be\u4ee5\u8fbe\u5230\u7535\u5f71\u7ea7\u6750\u8d28\u548c\u5149\u7167\u6548\u679c<\/td>\u6e38\u620f\u6e32\u67d3\u7684\u989c\u8272\u5e27\u3001\u8fd0\u52a8\u77e2\u91cf<\/td>\u91cd\u65b0\u589e\u5f3a\u5149\u7167\u3001\u6750\u8d28\u4e0e\u89d2\u8272\u5916\u89c2<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
    \n
  • 1. DLSS Super Resolution<\/li>\n<\/ul>\n\n\n\n

    \u8fd9\u9879\u6280\u672f\u4e3b\u8981\u662f\u7528\u6765\u548c\u4f20\u7edfAA\u5bf9\u6bd4\uff0c\u5982\u6700\u5e7f\u6cdb\u4f7f\u7528\u7684TAA\u3002\u5b9e\u9645\u4e0aTAA\u662f\u901a\u8fc7\u91c7\u6837\u5386\u53f2\u5e27\uff0c\u6765\u6269\u5927\u91c7\u6837\u51cf\u5c11\u952f\u9f7f\u3002\u800cDLSS\u8fd9\u91cc\u7528\u4e86\u4e00\u4e2aTransformer\u6a21\u578b\uff0c\u6765\u9884\u6d4b\u6a21\u578b\u7684\u8d85\u5206\u7f3a\u5931\u4fe1\u606f\uff08\u7c7b\u4f3c\u751f\u56fe\u7684\u5957\u8def\uff0c\u53ea\u662f\u4fe1\u606f\u66f4\u5168\u9762\uff09<\/p>\n\n\n\n

      \n
    • 2. DLSS Ray Reconstruction<\/li>\n<\/ul>\n\n\n\n

      \u8fd9\u4e00\u90e8\u5206\u7684\u5e94\u7528\u573a\u666f\u662f\u4f20\u7edfpath tracing\uff0c\u800c\u4e0d\u662f\u4efb\u610fray tracing\u3002\u5b9e\u9645\u66f4\u50cf\u662f\u4e00\u4e2a\u9ad8\u6548\u7684neural denoiser\u3002\u5f53\u7136\uff0c\u4eceAMD\u76f8\u5173\u7684\u6280\u672f\u8fdb\u5c55\u6765\u770b\uff0c\u795e\u7ecf\u7f51\u7edc\u65b9\u6cd5\u5f88\u6709\u5e0c\u671b\u5e94\u7528\u5728radiance cache\u4e0a\uff0c\u4ece\u800c\u63a5\u8fd1RTPT\u3002<\/p>\n\n\n\n

        \n
      • 3 DLSS Frame Generation<\/li>\n<\/ul>\n\n\n\n

        \u611f\u89c9\u4e0a\u548c\u8d85\u5206\u5dee\u4e0d\u591a\u3002DLSS 4.5\u5ba3\u79f0\u80fd\u591f\u8fbe\u5230\u6700\u9ad8 render 1 gen 5\uff0c\u4f46\u4ece\u611f\u53d7\u4e0a\uff0c\u8fd9\u4e2afeature\u5e94\u8be5\u6bd4\u8d85\u5206\u66f4\u4e0d\u9002\u5e94\u9ad8\u901f\u8fd0\u52a8\u573a\u666f\uff0c\u9664\u975e\u5176\u5728\u5904\u7406\u5f53\u524d\u5e27\u7684\u65f6\u5019\uff0c\u80fd\u591f\u770b\u5230\u548cTAA\u7c7b\u4f3c\u7684\u4fe1\u606f\uff08\u4f46\u8fd9\u6837\u8f93\u5165\u5c31\u592a\u591a\u4e86\u5427\uff0c\u800c\u4e14\u6781\u5ea6\u4f9d\u8d56\u5b9a\u5236\u5316\uff09\u3002\u5149\u6d41\/\u8fd0\u52a8\u4fe1\u606f\u662f\uff1f<\/p>\n\n\n\n

          \n
        • 4. DLSS 5<\/li>\n<\/ul>\n\n\n\n

          \u8fd9\u4e2a\u6280\u672f\u770b\u7740\u66f4\u62bd\u8c61\u4e86\uff0c\u5ba3\u4f20\u4e0a\u7684\u8bf4\u6cd5\u7c7b\u4f3c\uff0c\u7406\u89e3PBR\u6750\u8d28\uff0c\u7c7b\u4f3c\u5934\u53d1\uff0c\u5e03\u6599\uff0c\u76ae\u80a4\u8fd9\u4e9b\u4e0d\u597d\u5904\u7406\u7684\u90e8\u5206\uff0c\u8fdb\u884c\u91cd\u5149\u7167\u3002\u66f4\u50cf\u662f\u751f\u6210\u6a21\u578b\u7684\u8def\u5f84\uff0c\u4e5f\u5c31\u662f\u5982\u679c\u8981\u5b9e\u9645\u5e94\u7528\u7684\u8bdd\uff0c\u5927\u6982\u7387\u9700\u8981\u5bf9\u5e94\u7684fine tune\uff0c\u4e0d\u7136\u76f8\u5f53\u4e8e\u4e3a\u7f8e\u672f\u98ce\u683c\u5404\u5f02\u7684\u753b\u9762\u52a0\u4e0a\u4e86\u4e00\u5c42\u4e00\u6837\u7684\u6ee4\u955c\uff1f\u5728\u5361\u6e32\u4e0a\u66f4\u4e0d\u73b0\u5b9e\u3002\u6027\u80fd\u4e0a\uff0c\u636e\u8bf4\u5c55\u793a\u4f1a\u4e0a\u4e13\u95e8\u7528\u4e00\u5f205090\u6765\u8fd0\u884c\u6a21\u578b\uff0c\u663e\u7136\u8ddd\u79bb\u751f\u4ea7\u5e94\u7528\u8fd8\u6709\u8ddd\u79bb\u3002<\/p>\n\n\n\n

          2. DLSS 2 \u7ec6\u8282\u62c6\u89e3<\/h2>\n\n\n\n

          DLSS 2 \u5185\u90e8\u4ec5\u63d0\u4f9b\u4e00\u4e2a\u8d85\u5206\u7684\u529f\u80fd\uff0c\u56e0\u4e3a\u6709\u70b9\u8001\uff0c\u6240\u4ee5\u4ec5\u7528\u4f5c\u6280\u672f\u53c2\u8003\u3002<\/p>\n\n\n\n

          DLSS 2 \u7684\u672c\u4f53\u662f\u4e00\u4e2aCNN\uff0c\u75315\u5c42encode\u548c5\u5c42decode\u6784\u6210\uff0c\u53c2\u6570\u603b\u548c\u7ea61M\u3002<\/p>\n\n\n\n

          \u5176\u4e2d\uff0c\u63a5\u6536\u4fe1\u606f\u7684\u5f62\u5f0f\u6709\u70b9\u50cfTAA\uff0cdlss\u5185\u90e8\u4f1a\u7ef4\u62a4\u4e00\u4e2a\u5386\u53f2\u5e27\uff0c\u5305\u542b\u9ad8\u5206\u8fa8\u7387<\/strong>\u7684\u4e00\u5f20\u5386\u53f2\u56fe\u50cf<\/p>\n\n\n\n

          \u751f\u6210\u65b0\u5e27\u7684\u65b9\u5f0f\uff1a<\/strong><\/p>\n\n\n\n

          DLSS \u7684 CNN \u4e0d\u76f4\u63a5\u8f93\u51fa\u6700\u7ec8 RGB\u3002\u5b83\u8f93\u51fa\u4e09\u4e2a\u5404\u5411\u5f02\u6027\u9ad8\u65af\u6838\u5f62\u72b6\u53c2\u6570<\/strong> z0\u3001z1\u3001z2<\/code>\uff0c\u4e94\u4e2a\u4e0b\u4e00\u5e27\u9690\u85cf\u72b6\u6001\u53c2\u6570\uff0c\u4ee5\u53ca\u4e24\u4e2a\u65f6\u7a7a\u6df7\u5408\u53c2\u6570<\/strong> z8\u3001z9<\/code>\u3002<\/p>\n\n\n\n

            \n
          • z0\u3001z1\u3001z2<\/code>\uff1a\u51b3\u5b9a\u5f53\u524d\u5e27\u7a7a\u95f4\u6ee4\u6ce2\u6838\u7684\u5bbd\u5ea6\u3001\u5f62\u72b6\u548c\u65b9\u5411\u3002<\/li>\n\n\n\n
          • z8<\/code>\uff1a\u63a7\u5236\u5f53\u524d\u5e27\u989c\u8272\u8fdb\u5165\u5386\u53f2\u7f13\u51b2\u533a\u7684\u901f\u5ea6\uff0c\u5373 Temporal \u7d2f\u8ba1\u5f3a\u5ea6\u3002<\/li>\n\n\n\n
          • z9<\/code>\uff1a\u63a7\u5236\u7a7a\u95f4\u6ee4\u6ce2\u7ed3\u679c\u4e0e\u65f6\u57df\u7d2f\u8ba1\u7ed3\u679c\u4e4b\u95f4\u7684\u6700\u7ec8\u6df7\u5408\u3002<\/li>\n<\/ul>\n\n\n\n

            1. \u5c06\u9ad8\u5206\u8fa8\u7387\u8f93\u51fa\u50cf\u7d20\u6620\u5c04\u5230\u4f4e\u5206\u8fa8\u7387\u56fe\u50cf<\/h3>\n\n\n\n

            \u5bf9\u4e8e\u6bcf\u4e00\u4e2a\u9ad8\u5206\u8fa8\u7387\u8f93\u51fa\u50cf\u7d20 p=(x,y)<\/code>\uff0c\u9996\u5148\u8ba1\u7b97\u5b83\u5728\u4f4e\u5206\u8fa8\u7387\u8f93\u5165\u56fe\u50cf\u4e2d\u7684\u8fde\u7eed\u4e9a\u50cf\u7d20\u4f4d\u7f6e\uff1a<\/p>\n\n\n\n

            q<\/mi>=<\/mo>(<\/mo>p<\/mi>+<\/mo>0.5<\/mn>)<\/mo>\u2299<\/mo>S<\/mi><\/mtext>in<\/mi><\/mrow><\/msub>S<\/mi><\/mtext>out<\/mi><\/mrow><\/msub><\/mfrac>\u2212<\/mo>0.5<\/mn>+<\/mo>J<\/mi><\/mrow>q=(p+0.5)\\odot\\frac{S_{\\mathrm{in}}}{S_{\\mathrm{out}}}-0.5+J<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u5176\u4e2d\uff0cJ<\/code> \u662f\u5f53\u524d\u5e27\u7684 Jitter\uff0c\u2299<\/code> \u8868\u793a\u5bbd\u5ea6\u548c\u9ad8\u5ea6\u5206\u522b\u76f8\u4e58\u3002\u63a5\u7740\u627e\u5230\u6700\u8fd1\u7684\u4f4e\u5206\u8fa8\u7387\u50cf\u7d20\u4e2d\u5fc3\uff0c\u5e76\u8ba1\u7b97\u4e9a\u50cf\u7d20\u504f\u79fb\uff1a<\/p>\n\n\n\n

            n<\/mi>=<\/mo>round<\/mi>\u2061<\/mo>(<\/mo>q<\/mi>)<\/mo>,<\/mo><\/mspace>\u03b4<\/mi>=<\/mo>n<\/mi>\u2212<\/mo>q<\/mi><\/mrow>n=\\operatorname{round}(q),\\qquad \\delta=n-q<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u4f8b\u5982\uff0c\u8f93\u5165\u56fe\u50cf\u4e3a 4\u00d74<\/code>\u3001\u8f93\u51fa\u56fe\u50cf\u4e3a 6\u00d76<\/code>\uff0c\u8ba1\u7b97\u8f93\u51fa\u50cf\u7d20 p=(2,2)<\/code>\uff0c\u5e76\u5047\u8bbe Jitter \u4e3a 0\uff1a<\/p>\n\n\n\n

            q<\/mi>x<\/mi><\/msub>=<\/mo>q<\/mi>y<\/mi><\/msub>=<\/mo>(<\/mo>2<\/mn>+<\/mo>0.5<\/mn>)<\/mo>\u00d7<\/mo>4<\/mn>6<\/mn><\/mfrac>\u2212<\/mo>0.5<\/mn>=<\/mo>1.1667<\/mn><\/mrow>q_x=q_y=(2+0.5)\\times\\frac{4}{6}-0.5=1.1667<\/annotation><\/semantics><\/math><\/div>\n\n\n\n
            n<\/mi>=<\/mo>(<\/mo>1,1<\/mn>)<\/mo>,<\/mo><\/mspace>\u03b4<\/mi>=<\/mo>(<\/mo>\u2212<\/mo>0.1667<\/mn>,<\/mo>\u2212<\/mo>0.1667<\/mn>)<\/mo><\/mrow>n=(1,1),\\qquad \\delta=(-0.1667,-0.1667)<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u56e0\u6b64\uff0c\u8be5\u9ad8\u5206\u8fa8\u7387\u8f93\u51fa\u50cf\u7d20\u5bf9\u5e94\u4f4e\u5206\u8fa8\u7387\u56fe\u50cf\u4e2d\u7684\u8fde\u7eed\u4f4d\u7f6e (1.1667,1.1667)<\/code>\uff0c\u6700\u8fd1\u7684\u4f4e\u5206\u8fa8\u7387\u50cf\u7d20\u4e2d\u5fc3\u662f (1,1)<\/code>\u3002<\/p>\n\n\n\n

            2. \u51b3\u5b9a\u4f7f\u7528 3\u00d73 \u8fd8\u662f 5\u00d75 \u90bb\u57df<\/h3>\n\n\n\n

            3\u00d73 \u6216 5\u00d75 \u4e0d\u662f CNN \u5bf9\u6bcf\u4e2a\u50cf\u7d20\u5355\u72ec\u9009\u62e9\u7684\uff0c\u800c\u662f\u5728\u521b\u5efa DLSS Feature \u65f6\u6839\u636e\u521d\u59cb\u8f93\u5165\/\u8f93\u51fa\u6bd4\u4f8b\u5168\u5c40\u786e\u5b9a\uff1a<\/p>\n\n\n\n

            r<\/mi><\/mtext>scale<\/mi><\/mrow><\/msub>=<\/mo>W<\/mi><\/mtext>input<\/mi><\/mrow><\/msub>W<\/mi><\/mtext>output<\/mi><\/mrow><\/msub><\/mfrac><\/mrow>r_{\\mathrm{scale}}=\\frac{W_{\\mathrm{input}}}{W_{\\mathrm{output}}}<\/annotation><\/semantics><\/math><\/div>\n\n\n\n
            r<\/mi><\/mtext>scale<\/mi><\/mrow><\/msub><<\/mo>0.4<\/mn>\u21d2<\/mo>5<\/mn>\u00d7<\/mo>5<\/mn>,<\/mo><\/mspace>r<\/mi><\/mtext>scale<\/mi><\/mrow><\/msub>\u2265<\/mo>0.4<\/mn>\u21d2<\/mo>3<\/mn>\u00d7<\/mo>3<\/mn><\/mrow>r_{\\mathrm{scale}}<0.4\\Rightarrow 5\\times5,\\qquad r_{\\mathrm{scale}}\\geq0.4\\Rightarrow3\\times3<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u5f53\u5bbd\u5ea6\u6bd4\u5c0f\u4e8e 0.4 \u65f6\uff0c\u8f93\u5165\u5206\u8fa8\u7387\u5f88\u4f4e\uff0c3\u00d73 \u90bb\u57df\u80fd\u591f\u63d0\u4f9b\u7684\u6837\u672c\u4e0d\u8db3\uff0c\u56e0\u6b64\u4f7f\u7528 5\u00d75 \u83b7\u53d6\u66f4\u5927\u8303\u56f4\u7684\u4fe1\u606f\u3002\u5426\u5219\u9ed8\u8ba4\u4f7f\u7528 3\u00d73\u3002<\/p>\n\n\n\n

            \u8fd9\u9879\u9009\u62e9\u5bf9\u6574\u4e2a DLSS \u5b9e\u4f8b\u7edf\u4e00\u751f\u6548\u3002CNN \u52a8\u6001\u9884\u6d4b\u7684\u662f\u9ad8\u65af\u6838\u7684\u5f62\u72b6\u548c\u65b9\u5411\uff0c\u800c\u4e0d\u662f\u6bcf\u4e2a\u50cf\u7d20\u4f7f\u7528 3\u00d73 \u8fd8\u662f 5\u00d75\u3002<\/p>\n\n\n\n

            3. \u7528 CNN \u53c2\u6570\u6784\u9020\u5404\u5411\u5f02\u6027\u9ad8\u65af\u6838<\/h3>\n\n\n\n

            CNN \u8f93\u51fa\u7684 z0\u3001z1\u3001z2<\/code> \u9996\u5148\u88ab\u8f6c\u6362\u4e3a\u4e09\u4e2a\u534f\u65b9\u5dee\u53c2\u6570\uff1a<\/p>\n\n\n\n

            a<\/mi>=<\/mo>(<\/mo>1.7<\/mn><\/mspace>\u03c3<\/mi>(<\/mo>z<\/mi>0<\/mn><\/msub>)<\/mo>+<\/mo>0.17<\/mn>)<\/mo><\/mrow>2<\/mn><\/msup><\/mrow>a=\\left(1.7\\,\\sigma(z_0)+0.17\\right)^2<\/annotation><\/semantics><\/math><\/div>\n\n\n\n
            b<\/mi>=<\/mo>(<\/mo>1.7<\/mn><\/mspace>\u03c3<\/mi>(<\/mo>z<\/mi>1<\/mn><\/msub>)<\/mo>+<\/mo>0.17<\/mn>)<\/mo><\/mrow>2<\/mn><\/msup><\/mrow>b=\\left(1.7\\,\\sigma(z_1)+0.17\\right)^2<\/annotation><\/semantics><\/math><\/div>\n\n\n\n
            c<\/mi>=<\/mo>a<\/mi>b<\/mi><\/mrow><\/msqrt>\u22c5<\/mo>0.95<\/mn><\/mspace>tanh<\/mi>\u2061<\/mo><\/mrow>(<\/mo>z<\/mi>2<\/mn><\/msub>)<\/mo><\/mrow>c=\\sqrt{ab}\\cdot0.95\\tanh(z_2)<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u5176\u4e2d\uff0cSigmoid \u51fd\u6570\u4e3a\uff1a<\/p>\n\n\n\n

            \u03c3<\/mi>(<\/mo>z<\/mi>)<\/mo>=<\/mo>1<\/mn>1<\/mn>+<\/mo>e<\/mi>\u2212<\/mo>z<\/mi><\/mrow><\/msup><\/mrow><\/mfrac><\/mrow>\\sigma(z)=\\frac{1}{1+e^{-z}}<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u7136\u540e\u6784\u9020\u4e8c\u7ef4\u534f\u65b9\u5dee\u77e9\u9635\uff1a<\/p>\n\n\n\n

            \u03a3<\/mi><\/mrow>=<\/mo>[<\/mo>a<\/mi><\/mtd>c<\/mi><\/mtd><\/mtr>c<\/mi><\/mtd>b<\/mi><\/mtd><\/mtr><\/mtable>]<\/mo><\/mrow><\/mrow>\\Sigma=\\begin{bmatrix}a&c\\\\c&b\\end{bmatrix}<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            a<\/code> \u548c b<\/code> \u63a7\u5236\u9ad8\u65af\u6838\u5728\u4e24\u4e2a\u65b9\u5411\u4e0a\u7684\u5bbd\u5ea6\uff0cc<\/code> \u63a7\u5236\u9ad8\u65af\u6838\u7684\u65cb\u8f6c\u548c\u65b9\u5411\u76f8\u5173\u6027\u3002<\/p>\n\n\n\n

            \u5047\u8bbe CNN \u4e3a\u5f53\u524d\u533a\u57df\u8f93\u51fa\uff1a<\/p>\n\n\n\n

            z<\/mi>0<\/mn><\/msub>=<\/mo>z<\/mi>1<\/mn><\/msub>=<\/mo>z<\/mi>2<\/mn><\/msub>=<\/mo>0<\/mn><\/mrow>z_0=z_1=z_2=0<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u7531\u4e8e sigmoid(0)=0.5<\/code> \u4e14 tanh(0)=0<\/code>\uff0c\u53ef\u5f97\uff1a<\/p>\n\n\n\n

            a<\/mi>=<\/mo>b<\/mi>=<\/mo>(<\/mo>1.7<\/mn>\u00d7<\/mo>0.5<\/mn>+<\/mo>0.17<\/mn>)<\/mo>2<\/mn><\/msup>=<\/mo>1.0404<\/mn>,<\/mo><\/mspace>c<\/mi>=<\/mo>0<\/mn><\/mrow>a=b=(1.7\\times0.5+0.17)^2=1.0404,\\qquad c=0<\/annotation><\/semantics><\/math><\/div>\n\n\n\n
            \u03a3<\/mi><\/mrow>=<\/mo>[<\/mo>1.0404<\/mn><\/mtd>0<\/mn><\/mtd><\/mtr>0<\/mn><\/mtd>1.0404<\/mn><\/mtd><\/mtr><\/mtable>]<\/mo><\/mrow>,<\/mo><\/mspace>\u03a3<\/mi><\/mrow>\u2212<\/mo>1<\/mn><\/mrow><\/msup>=<\/mo>[<\/mo>0.96117<\/mn><\/mtd>0<\/mn><\/mtd><\/mtr>0<\/mn><\/mtd>0.96117<\/mn><\/mtd><\/mtr><\/mtable>]<\/mo><\/mrow><\/mrow>\\Sigma=\\begin{bmatrix}1.0404&0\\\\0&1.0404\\end{bmatrix},\\qquad \\Sigma^{-1}=\\begin{bmatrix}0.96117&0\\\\0&0.96117\\end{bmatrix}<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u6b64\u65f6\u5f97\u5230\u7684\u662f\u4e00\u4e2a\u6ca1\u6709\u65cb\u8f6c\u3001\u63a5\u8fd1\u5706\u5f62\u7684\u9ad8\u65af\u6838\u3002\u5982\u679c a\u2260b<\/code>\uff0c\u9ad8\u65af\u6838\u4f1a\u6cbf\u67d0\u4e00\u65b9\u5411\u62c9\u957f\uff1b\u5982\u679c c\u22600<\/code>\uff0c\u9ad8\u65af\u6838\u4f1a\u53d1\u751f\u65cb\u8f6c\uff0c\u4ece\u800c\u6cbf\u7740\u56fe\u50cf\u8fb9\u7f18\u65b9\u5411\u8fdb\u884c\u6ee4\u6ce2\u3002<\/p>\n\n\n\n

            4. \u8ba1\u7b97 3\u00d73 \u90bb\u57df\u7684\u9ad8\u65af\u6743\u91cd<\/h3>\n\n\n\n

            \u5bf9\u4f4e\u5206\u8fa8\u7387\u50cf\u7d20\u4e2d\u5fc3\u5468\u56f4\u7684\u6bcf\u4e2a\u90bb\u57df\u4f4d\u7f6e (i,j)<\/code>\uff0c\u5176\u4e2d 3\u00d73 \u6a21\u5f0f\u4e0b i,j\u2208{-1,0,1}<\/code>\uff0c\u5148\u8ba1\u7b97\u8003\u8651\u4e9a\u50cf\u7d20\u504f\u79fb\u540e\u7684\u8ddd\u79bb\uff1a<\/p>\n\n\n\n

            d<\/mi>i<\/mi>j<\/mi><\/mrow><\/msub>=<\/mo>[<\/mo>i<\/mi>+<\/mo>\u03b4<\/mi>x<\/mi><\/msub><\/mrow><\/mtd><\/mtr>j<\/mi>+<\/mo>\u03b4<\/mi>y<\/mi><\/msub><\/mrow><\/mtd><\/mtr><\/mtable>]<\/mo><\/mrow><\/mrow>d_{ij}=\\begin{bmatrix}i+\\delta_x\\\\j+\\delta_y\\end{bmatrix}<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u7136\u540e\u8ba1\u7b97\u672a\u5f52\u4e00\u5316\u9ad8\u65af\u6743\u91cd\uff1a<\/p>\n\n\n\n

            w<\/mi>~<\/mo><\/mover>i<\/mi>j<\/mi><\/mrow><\/msub>=<\/mo>exp<\/mi>\u2061<\/mo><\/mrow>(<\/mo>\u2212<\/mo>1<\/mn>2<\/mn><\/mfrac>d<\/mi>i<\/mi>j<\/mi><\/mrow>T<\/mi><\/msubsup>\u03a3<\/mi><\/mrow>\u2212<\/mo>1<\/mn><\/mrow><\/msup>d<\/mi>i<\/mi>j<\/mi><\/mrow><\/msub>)<\/mo><\/mrow><\/mrow>\\widetilde{w}_{ij}=\\exp\\left(-\\frac{1}{2}d_{ij}^{T}\\Sigma^{-1}d_{ij}\\right)<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u518d\u5c06\u6240\u6709\u90bb\u57df\u6743\u91cd\u5f52\u4e00\u5316\uff1a<\/p>\n\n\n\n

            w<\/mi>i<\/mi>j<\/mi><\/mrow><\/msub>=<\/mo>w<\/mi>~<\/mo><\/mover>i<\/mi>j<\/mi><\/mrow><\/msub>\u2211<\/mo>m<\/mi>=<\/mo>\u2212<\/mo>r<\/mi><\/mrow>r<\/mi><\/munderover><\/mrow>\u2211<\/mo>n<\/mi>=<\/mo>\u2212<\/mo>r<\/mi><\/mrow>r<\/mi><\/munderover><\/mrow>w<\/mi>~<\/mo><\/mover>m<\/mi>n<\/mi><\/mrow><\/msub><\/mstyle><\/mfrac><\/mrow>w_{ij}=\\frac{\\widetilde{w}_{ij}}{\\displaystyle\\sum_{m=-r}^{r}\\sum_{n=-r}^{r}\\widetilde{w}_{mn}}<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u5c06\u524d\u9762\u7684 \u03b4=(-0.1667,-0.1667)<\/code> \u548c\u534f\u65b9\u5dee\u77e9\u9635\u4ee3\u5165\u540e\uff0c\u5f97\u5230\u4ee5\u4e0b 3\u00d73 \u5f52\u4e00\u5316\u6743\u91cd\u77e9\u9635\uff1a<\/p>\n\n\n\n

            W<\/mi>=<\/mo>[<\/mo>0.05470<\/mn><\/mtd>0.10382<\/mn><\/mtd>0.07536<\/mn><\/mtd><\/mtr>0.10382<\/mn><\/mtd>0.19705<\/mn><\/mtd>0.14303<\/mn><\/mtd><\/mtr>0.07536<\/mn><\/mtd>0.14303<\/mn><\/mtd>0.10382<\/mn><\/mtd><\/mtr><\/mtable>]<\/mo><\/mrow><\/mrow>W=\\begin{bmatrix}0.05470&0.10382&0.07536\\\\0.10382&0.19705&0.14303\\\\0.07536&0.14303&0.10382\\end{bmatrix}<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u56e0\u4e3a\u76ee\u6807\u4e9a\u50cf\u7d20\u4f4d\u7f6e\u4f4d\u4e8e\u4e2d\u5fc3\u50cf\u7d20\u7684\u53f3\u4e0b\u65b9\uff0c\u6240\u4ee5\u53f3\u4fa7\u548c\u4e0b\u4fa7\u7684\u50cf\u7d20\u6743\u91cd\u76f8\u5bf9\u8f83\u5927\u3002<\/p>\n\n\n\n

            5. \u8ba1\u7b97\u5f53\u524d\u5e27\u7684\u7a7a\u95f4\u91cd\u5efa\u7ed3\u679c<\/h3>\n\n\n\n

            \u5047\u8bbe\u4f4e\u5206\u8fa8\u7387\u50cf\u7d20\u4e2d\u5fc3\u5468\u56f4\u7684 3\u00d73 \u7070\u5ea6\u503c\u4e3a\uff1a<\/p>\n\n\n\n

            C<\/mi>=<\/mo>[<\/mo>0.10<\/mn><\/mtd>0.10<\/mn><\/mtd>0.20<\/mn><\/mtd><\/mtr>0.10<\/mn><\/mtd>0.40<\/mn><\/mtd>0.80<\/mn><\/mtd><\/mtr>0.20<\/mn><\/mtd>0.80<\/mn><\/mtd>1.00<\/mn><\/mtd><\/mtr><\/mtable>]<\/mo><\/mrow><\/mrow>C=\\begin{bmatrix}0.10&0.10&0.20\\\\0.10&0.40&0.80\\\\0.20&0.80&1.00\\end{bmatrix}<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u4f7f\u7528\u9ad8\u65af\u6743\u91cd\u5bf9\u90bb\u57df\u989c\u8272\u52a0\u6743\u6c42\u548c\uff1a<\/p>\n\n\n\n

            C<\/mi><\/mtext>filter<\/mi><\/mrow><\/msub>=<\/mo>\u2211<\/mo>i<\/mi>=<\/mo>\u2212<\/mo>r<\/mi><\/mrow>r<\/mi><\/munderover><\/mrow>\u2211<\/mo>j<\/mi>=<\/mo>\u2212<\/mo>r<\/mi><\/mrow>r<\/mi><\/munderover><\/mrow>w<\/mi>i<\/mi>j<\/mi><\/mrow><\/msub><\/mspace>C<\/mi>(<\/mo>n<\/mi>x<\/mi><\/msub>+<\/mo>i<\/mi>,<\/mo>n<\/mi>y<\/mi><\/msub>+<\/mo>j<\/mi>)<\/mo><\/mrow>C_{\\mathrm{filter}}=\\sum_{i=-r}^{r}\\sum_{j=-r}^{r}w_{ij}\\,C(n_x+i,n_y+j)<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u4ee3\u5165\u4f8b\u5b50\u4e2d\u7684\u6570\u503c\uff1a<\/p>\n\n\n\n

            C<\/mi><\/mtext>filter<\/mi><\/mrow><\/msub>=<\/mo><\/mrow><\/mrow><\/mtd>0.05470<\/mn>\u00d7<\/mo>0.10<\/mn>+<\/mo>0.10382<\/mn>\u00d7<\/mo>0.10<\/mn>+<\/mo>0.07536<\/mn>\u00d7<\/mo>0.20<\/mn><\/mrow><\/mtd><\/mtr><\/mrow><\/mtd>+<\/mo>0.10382<\/mn>\u00d7<\/mo>0.10<\/mn>+<\/mo>0.19705<\/mn>\u00d7<\/mo>0.40<\/mn>+<\/mo>0.14303<\/mn>\u00d7<\/mo>0.80<\/mn><\/mrow><\/mtd><\/mtr><\/mrow><\/mtd>+<\/mo>0.07536<\/mn>\u00d7<\/mo>0.20<\/mn>+<\/mo>0.14303<\/mn>\u00d7<\/mo>0.80<\/mn>+<\/mo>0.10382<\/mn>\u00d7<\/mo>1.00<\/mn><\/mrow><\/mtd><\/mtr>\u2248<\/mo><\/mrow><\/mrow><\/mtd>0.46787<\/mn><\/mtd><\/mtr><\/mtable>\\begin{aligned}C_{\\mathrm{filter}}={}&0.05470\\times0.10+0.10382\\times0.10+0.07536\\times0.20\\\\&+0.10382\\times0.10+0.19705\\times0.40+0.14303\\times0.80\\\\&+0.07536\\times0.20+0.14303\\times0.80+0.10382\\times1.00\\\\\\approx{}&0.46787\\end{aligned}<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u56e0\u6b64\uff0c\u8be5\u8f93\u51fa\u50cf\u7d20\u4ec5\u6839\u636e\u5f53\u524d\u4f4e\u5206\u8fa8\u7387\u56fe\u50cf\u5f97\u5230\u7684\u7a7a\u95f4\u91cd\u5efa\u7ed3\u679c\u4e3a 0.46787<\/code>\u3002<\/p>\n\n\n\n

            \u771f\u5b9e\u56fe\u50cf\u662f RGB \u56fe\u50cf\u65f6\uff0c\u540c\u4e00\u7ec4\u7a7a\u95f4\u6743\u91cd\u5206\u522b\u4f5c\u7528\u4e8e\u4e09\u4e2a\u989c\u8272\u901a\u9053\uff1a<\/p>\n\n\n\n

            R<\/mi><\/mtext>filter<\/mi><\/mrow><\/msub>=<\/mo>\u2211<\/mo>w<\/mi>i<\/mi>j<\/mi><\/mrow><\/msub>R<\/mi>i<\/mi>j<\/mi><\/mrow><\/msub>,<\/mo><\/mspace>G<\/mi><\/mtext>filter<\/mi><\/mrow><\/msub>=<\/mo>\u2211<\/mo>w<\/mi>i<\/mi>j<\/mi><\/mrow><\/msub>G<\/mi>i<\/mi>j<\/mi><\/mrow><\/msub>,<\/mo><\/mspace>B<\/mi><\/mtext>filter<\/mi><\/mrow><\/msub>=<\/mo>\u2211<\/mo>w<\/mi>i<\/mi>j<\/mi><\/mrow><\/msub>B<\/mi>i<\/mi>j<\/mi><\/mrow><\/msub><\/mrow>R_{\\mathrm{filter}}=\\sum w_{ij}R_{ij},\\qquad G_{\\mathrm{filter}}=\\sum w_{ij}G_{ij},\\qquad B_{\\mathrm{filter}}=\\sum w_{ij}B_{ij}<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            6. Temporal\uff1a\u5c06\u5f53\u524d\u5e27\u52a0\u5165\u9ad8\u5206\u8fa8\u7387\u5386\u53f2<\/h3>\n\n\n\n

            z8<\/code> \u63a7\u5236\u5f53\u524d\u4f4e\u5206\u8fa8\u7387\u50cf\u7d20\u4e2d\u5fc3\u989c\u8272\u8fdb\u5165\u5386\u53f2\u7f13\u51b2\u533a\u7684\u901f\u5ea6\u3002\u9996\u5148\u8ba1\u7b97\uff1a<\/p>\n\n\n\n

            g<\/mi>j<\/mi><\/msub>=<\/mo>\u03c3<\/mi>(<\/mo>z<\/mi>8<\/mn><\/msub>)<\/mo><\/mrow>g_j=\\sigma(z_8)<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u5b9e\u9645\u7684\u5f53\u524d\u5e27\u5386\u53f2\u66f4\u65b0\u7387\u4e3a\uff1a<\/p>\n\n\n\n

            \u03b1<\/mi>=<\/mo>\u03b1<\/mi><\/mtext>learned<\/mi><\/mrow><\/msub>\u22c5<\/mo>W<\/mi><\/mtext>out<\/mi><\/mrow><\/msub>W<\/mi><\/mtext>in<\/mi><\/mrow><\/msub><\/mfrac>\u22c5<\/mo>H<\/mi><\/mtext>out<\/mi><\/mrow><\/msub>H<\/mi><\/mtext>in<\/mi><\/mrow><\/msub><\/mfrac>\u22c5<\/mo>g<\/mi>j<\/mi><\/msub><\/mrow>\\alpha=\\alpha_{\\mathrm{learned}}\\cdot\\frac{W_{\\mathrm{out}}}{W_{\\mathrm{in}}}\\cdot\\frac{H_{\\mathrm{out}}}{H_{\\mathrm{in}}}\\cdot g_j<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u5176\u4e2d\uff0c\u03b1learned<\/code> \u662f\u6a21\u578b\u4fdd\u5b58\u7684\u57fa\u7840\u5386\u53f2\u878d\u5408\u7cfb\u6570\u3002\u5b83\u53ef\u4ee5\u7406\u89e3\u4e3a\u5386\u53f2\u66f4\u65b0\u7387<\/strong>\uff0c\u4f46\u4e0d\u662f\u8bad\u7ec3\u795e\u7ecf\u7f51\u7edc\u65f6\u4f18\u5316\u5668\u4f7f\u7528\u7684 learning rate\u3002<\/p>\n\n\n\n

            \u7ee7\u7eed\u4f7f\u7528\u524d\u9762\u7684 4\u00d74 \u2192 6\u00d76<\/code> \u4f8b\u5b50\u3002\u5047\u8bbe\uff1a<\/p>\n\n\n\n

            \u03b1<\/mi><\/mtext>learned<\/mi><\/mrow><\/msub>=<\/mo>0.035<\/mn>,<\/mo><\/mspace>z<\/mi>8<\/mn><\/msub>=<\/mo>0<\/mn>,<\/mo><\/mspace>g<\/mi>j<\/mi><\/msub>=<\/mo>\u03c3<\/mi>(<\/mo>0<\/mn>)<\/mo>=<\/mo>0.5<\/mn><\/mrow>\\alpha_{\\mathrm{learned}}=0.035,\\qquad z_8=0,\\qquad g_j=\\sigma(0)=0.5<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u5219\u5f53\u524d\u5e27\u7684\u5386\u53f2\u66f4\u65b0\u7387\u4e3a\uff1a<\/p>\n\n\n\n

            \u03b1<\/mi>=<\/mo>0.035<\/mn>\u00d7<\/mo>1.5<\/mn>\u00d7<\/mo>1.5<\/mn>\u00d7<\/mo>0.5<\/mn>=<\/mo>0.039375<\/mn><\/mrow>\\alpha=0.035\\times1.5\\times1.5\\times0.5=0.039375<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u5047\u8bbe\u5f53\u524d\u4f4e\u5206\u8fa8\u7387\u50cf\u7d20\u4e2d\u5fc3\u989c\u8272\u4e3a\uff1a<\/p>\n\n\n\n

            C<\/mi><\/mtext>current<\/mi><\/mrow><\/msub>=<\/mo>0.40<\/mn><\/mrow>C_{\\mathrm{current}}=0.40<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u8fd0\u52a8\u77e2\u91cf\u5bf9\u9f50\u540e\u7684\u9ad8\u5206\u8fa8\u7387\u5386\u53f2\u989c\u8272\u4e3a\uff1a<\/p>\n\n\n\n

            C<\/mi><\/mtext>history<\/mi><\/mrow><\/msub>=<\/mo>0.55<\/mn><\/mrow>C_{\\mathrm{history}}=0.55<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u65f6\u57df\u7d2f\u8ba1\u7ed3\u679c\u4e3a\uff1a<\/p>\n\n\n\n

            C<\/mi><\/mtext>temporal<\/mi><\/mrow><\/msub>=<\/mo>\u03b1<\/mi>C<\/mi><\/mtext>current<\/mi><\/mrow><\/msub>+<\/mo>(<\/mo>1<\/mn>\u2212<\/mo>\u03b1<\/mi>)<\/mo>C<\/mi><\/mtext>history<\/mi><\/mrow><\/msub><\/mrow>C_{\\mathrm{temporal}}=\\alpha C_{\\mathrm{current}}+(1-\\alpha)C_{\\mathrm{history}}<\/annotation><\/semantics><\/math><\/div>\n\n\n\n
            C<\/mi><\/mtext>temporal<\/mi><\/mrow><\/msub>=<\/mo>0.039375<\/mn>\u00d7<\/mo>0.40<\/mn>+<\/mo>0.960625<\/mn>\u00d7<\/mo>0.55<\/mn>\u2248<\/mo>0.54409<\/mn><\/mrow>C_{\\mathrm{temporal}}=0.039375\\times0.40+0.960625\\times0.55\\approx0.54409<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u56e0\u6b64\uff0c\u8fd9\u4e00\u5e27\u53ea\u7528\u7ea6 3.94% \u7684\u6bd4\u4f8b\u66f4\u65b0\u5386\u53f2\uff0c\u5269\u4f59\u7ea6 96.06% \u6765\u81ea\u5df2\u7ecf\u8fd0\u52a8\u8865\u507f\u7684\u5386\u53f2\u7ed3\u679c\u3002\u8fd9\u6837\u53ef\u4ee5\u7a33\u5b9a\u5730\u79ef\u7d2f\u4e0d\u540c\u5e27 Jitter \u5e26\u6765\u7684\u4e9a\u50cf\u7d20\u91c7\u6837\u4fe1\u606f\u3002<\/p>\n\n\n\n

            7. \u5386\u53f2\u5e27\u662f\u600e\u6837\u9010\u5e27\u7d2f\u8ba1\u7684<\/h3>\n\n\n\n

            \u5386\u53f2\u66f4\u65b0\u662f\u4e00\u4e2a\u9012\u5f52\u8fc7\u7a0b\uff1a<\/p>\n\n\n\n

            H<\/mi>t<\/mi><\/msub>=<\/mo>\u03b1<\/mi>C<\/mi>t<\/mi><\/msub>+<\/mo>(<\/mo>1<\/mn>\u2212<\/mo>\u03b1<\/mi>)<\/mo>H<\/mi>t<\/mi>\u2212<\/mo>1<\/mn><\/mrow><\/msub><\/mrow>H_t=\\alpha C_t+(1-\\alpha)H_{t-1}<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u5c06\u4e0a\u4e00\u5e27\u5386\u53f2\u7ee7\u7eed\u5c55\u5f00\uff1a<\/p>\n\n\n\n

            H<\/mi>t<\/mi><\/msub>=<\/mo>\u03b1<\/mi>C<\/mi>t<\/mi><\/msub>+<\/mo>\u03b1<\/mi>(<\/mo>1<\/mn>\u2212<\/mo>\u03b1<\/mi>)<\/mo>C<\/mi>t<\/mi>\u2212<\/mo>1<\/mn><\/mrow><\/msub>+<\/mo>\u03b1<\/mi>(<\/mo>1<\/mn>\u2212<\/mo>\u03b1<\/mi>)<\/mo>2<\/mn><\/msup>C<\/mi>t<\/mi>\u2212<\/mo>2<\/mn><\/mrow><\/msub>+<\/mo>\u22ef<\/mo><\/mrow>H_t=\\alpha C_t+\\alpha(1-\\alpha)C_{t-1}+\\alpha(1-\\alpha)^2C_{t-2}+\\cdots<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u56e0\u6b64\uff0c\u8ddd\u79bb\u5f53\u524d\u5e27 k<\/code> \u5e27\u7684\u989c\u8272\u8d21\u732e\u6743\u91cd\u4e3a\uff1a<\/p>\n\n\n\n

            w<\/mi>k<\/mi><\/msub>=<\/mo>\u03b1<\/mi>(<\/mo>1<\/mn>\u2212<\/mo>\u03b1<\/mi>)<\/mo>k<\/mi><\/msup><\/mrow>w_k=\\alpha(1-\\alpha)^k<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u5f53 \u03b1=0.039375<\/code> \u65f6\uff1a<\/p>\n\n\n\n

            w<\/mi>0<\/mn><\/msub>=<\/mo>0.039375<\/mn><\/mrow>w_0=0.039375<\/annotation><\/semantics><\/math><\/div>\n\n\n\n
            w<\/mi>1<\/mn><\/msub>=<\/mo>0.039375<\/mn>\u00d7<\/mo>0.960625<\/mn>\u2248<\/mo>0.03782<\/mn><\/mrow>w_1=0.039375\\times0.960625\\approx0.03782<\/annotation><\/semantics><\/math><\/div>\n\n\n\n
            w<\/mi>2<\/mn><\/msub>=<\/mo>0.039375<\/mn>\u00d7<\/mo>0.960625<\/mn>2<\/mn><\/msup>\u2248<\/mo>0.03633<\/mn><\/mrow>w_2=0.039375\\times0.960625^2\\approx0.03633<\/annotation><\/semantics><\/math><\/div>\n\n\n\n
            w<\/mi>10<\/mn><\/msub>=<\/mo>0.039375<\/mn>\u00d7<\/mo>0.960625<\/mn>10<\/mn><\/msup>\u2248<\/mo>0.02635<\/mn><\/mrow>w_{10}=0.039375\\times0.960625^{10}\\approx0.02635<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u867d\u7136\u8d8a\u65e9\u7684\u5355\u5e27\u6743\u91cd\u4f1a\u9010\u6e10\u8870\u51cf\uff0c\u4f46\u5f88\u591a\u5386\u53f2\u5e27\u7684\u8d21\u732e\u76f8\u52a0\u540e\u4ecd\u7136\u975e\u5e38\u660e\u663e\u3002\u6700\u8fd1 N<\/code> \u5e27\u7d2f\u8ba1\u8fdb\u5165\u5386\u53f2\u7684\u603b\u6743\u91cd\u4e3a\uff1a<\/p>\n\n\n\n

            A<\/mi>N<\/mi><\/msub>=<\/mo>1<\/mn>\u2212<\/mo>(<\/mo>1<\/mn>\u2212<\/mo>\u03b1<\/mi>)<\/mo>N<\/mi><\/msup><\/mrow>A_N=1-(1-\\alpha)^N<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u4ee3\u5165 \u03b1=0.039375<\/code>\uff1a<\/p>\n\n\n\n

            A<\/mi>10<\/mn><\/msub>=<\/mo>1<\/mn>\u2212<\/mo>0.960625<\/mn>10<\/mn><\/msup>\u2248<\/mo>33.1<\/mn>%<\/mi><\/mrow>A_{10}=1-0.960625^{10}\\approx33.1\\%<\/annotation><\/semantics><\/math><\/div>\n\n\n\n
            A<\/mi>30<\/mn><\/msub>=<\/mo>1<\/mn>\u2212<\/mo>0.960625<\/mn>30<\/mn><\/msup>\u2248<\/mo>70.1<\/mn>%<\/mi><\/mrow>A_{30}=1-0.960625^{30}\\approx70.1\\%<\/annotation><\/semantics><\/math><\/div>\n\n\n\n
            A<\/mi>60<\/mn><\/msub>=<\/mo>1<\/mn>\u2212<\/mo>0.960625<\/mn>60<\/mn><\/msup>\u2248<\/mo>91.0<\/mn>%<\/mi><\/mrow>A_{60}=1-0.960625^{60}\\approx91.0\\%<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u8fd9\u4f53\u73b0\u4e86\u65f6\u57df\u7d2f\u8ba1\u7684\u4f5c\u7528\uff1a\u5355\u5e27\u53ea\u52a0\u5165\u5c11\u91cf\u65b0\u4fe1\u606f\uff0c\u4f46\u7ecf\u8fc7\u51e0\u5341\u5e27\u4e0d\u540c Jitter \u4f4d\u7f6e\u7684\u91c7\u6837\u540e\uff0c\u5386\u53f2\u7f13\u51b2\u533a\u53ef\u4ee5\u9010\u6e10\u79ef\u7d2f\u66f4\u591a\u4e9a\u50cf\u7d20\u4fe1\u606f\u3002<\/p>\n\n\n\n

            \u5f53\u7269\u4f53\u8fd0\u52a8\u3001\u53d1\u751f\u53cd\u906e\u6321\u6216\u5386\u53f2\u989c\u8272\u4e0d\u518d\u53ef\u4fe1\u65f6\uff0cCNN \u7684\u95e8\u63a7\u4ee5\u53ca\u8fd0\u52a8\u77e2\u91cf\u5dee\u5f02\u3001\u989c\u8272\u5dee\u5f02\u548c\u6df1\u5ea6\u4fe1\u606f\u4f1a\u964d\u4f4e\u5386\u53f2\u7684\u53ef\u4fe1\u5ea6\uff0c\u4f7f\u7ed3\u679c\u66f4\u504f\u5411\u5f53\u524d\u5e27\u3002<\/p>\n\n\n\n

            8. Spatial\uff1a\u6df7\u5408\u7a7a\u95f4\u6ee4\u6ce2\u4e0e\u65f6\u57df\u7d2f\u8ba1\u7ed3\u679c<\/h3>\n\n\n\n

            z9<\/code> \u51b3\u5b9a\u7a7a\u95f4\u6ee4\u6ce2\u7ed3\u679c Cfilter<\/code> \u4e0e\u65f6\u57df\u7d2f\u8ba1\u7ed3\u679c Ctemporal<\/code> \u7684\u6700\u7ec8\u6df7\u5408\u6bd4\u4f8b\u3002\u9996\u5148\u8ba1\u7b97\u7a7a\u95f4\u95e8\u63a7\uff1a<\/p>\n\n\n\n

            g<\/mi>=<\/mo>clamp<\/mi>\u2061<\/mo>(<\/mo>1.25<\/mn>(<\/mo>\u03c3<\/mi>(<\/mo>z<\/mi>9<\/mn><\/msub>)<\/mo>\u2212<\/mo>0.5<\/mn>)<\/mo><\/mrow>+<\/mo>0.5<\/mn>,<\/mo> <\/mtext>g<\/mi>min<\/mi>\u2061<\/mo><\/mrow><\/msub>,<\/mo> <\/mtext>1<\/mn>)<\/mo><\/mrow><\/mrow>g=\\operatorname{clamp}\\left(1.25\\left(\\sigma(z_9)-0.5\\right)+0.5,\\ g_{\\min},\\ 1\\right)<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u6700\u7ec8\u8f93\u51fa\u989c\u8272\u4e3a\uff1a<\/p>\n\n\n\n

            C<\/mi><\/mtext>out<\/mi><\/mrow><\/msub>=<\/mo>(<\/mo>1<\/mn>\u2212<\/mo>g<\/mi>)<\/mo>C<\/mi><\/mtext>temporal<\/mi><\/mrow><\/msub>+<\/mo>g<\/mi>C<\/mi><\/mtext>filter<\/mi><\/mrow><\/msub><\/mrow>C_{\\mathrm{out}}=(1-g)C_{\\mathrm{temporal}}+gC_{\\mathrm{filter}}<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u5728\u5f53\u524d\u4f8b\u5b50\u4e2d\uff0c\u5047\u8bbe z9=0<\/code>\uff0c\u5219\uff1a<\/p>\n\n\n\n

            g<\/mi>=<\/mo>1.25<\/mn>(<\/mo>\u03c3<\/mi>(<\/mo>0<\/mn>)<\/mo>\u2212<\/mo>0.5<\/mn>)<\/mo><\/mrow>+<\/mo>0.5<\/mn>=<\/mo>0.5<\/mn><\/mrow>g=1.25\\left(\\sigma(0)-0.5\\right)+0.5=0.5<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u4ee3\u5165\u524d\u9762\u8ba1\u7b97\u51fa\u7684\u65f6\u57df\u7ed3\u679c\u548c\u7a7a\u95f4\u6ee4\u6ce2\u7ed3\u679c\uff1a<\/p>\n\n\n\n

            C<\/mi><\/mtext>out<\/mi><\/mrow><\/msub>=<\/mo>0.5<\/mn>\u00d7<\/mo>0.54409<\/mn>+<\/mo>0.5<\/mn>\u00d7<\/mo>0.46787<\/mn>\u2248<\/mo>0.50598<\/mn><\/mrow>C_{\\mathrm{out}}=0.5\\times0.54409+0.5\\times0.46787\\approx0.50598<\/annotation><\/semantics><\/math><\/div>\n\n\n\n

            \u6574\u4e2a\u4f8b\u5b50\u7684\u5173\u952e\u6570\u503c\u5982\u4e0b\uff1a<\/p>\n\n\n\n

              \n
            • \u4f4e\u5206\u8fa8\u7387\u4e2d\u5fc3\u989c\u8272\uff1a0.40000<\/code><\/li>\n\n\n\n
            • \u795e\u7ecf\u7a7a\u95f4\u6ee4\u6ce2\u7ed3\u679c\uff1a0.46787<\/code><\/li>\n\n\n\n
            • \u8fd0\u52a8\u8865\u507f\u5386\u53f2\u989c\u8272\uff1a0.55000<\/code><\/li>\n\n\n\n
            • \u65f6\u57df\u7d2f\u8ba1\u7ed3\u679c\uff1a0.54409<\/code><\/li>\n\n\n\n
            • \u6700\u7ec8\u8f93\u51fa\u989c\u8272\uff1a0.50598<\/code><\/li>\n<\/ul>\n\n\n\n

              z9<\/code> \u8d8a\u5927\uff0c\u6700\u7ec8\u7ed3\u679c\u8d8a\u76f8\u4fe1\u5f53\u524d\u5e27\u7684\u7a7a\u95f4\u6ee4\u6ce2\u91cd\u5efa\uff1bz9<\/code> \u8d8a\u5c0f\uff0c\u6700\u7ec8\u7ed3\u679c\u8d8a\u76f8\u4fe1\u7ecf\u8fc7\u8fd0\u52a8\u8865\u507f\u7684\u591a\u5e27\u5386\u53f2\u3002<\/p>\n\n\n\n

              \u603b\u7ed3\uff1a<\/strong>DLSS \u7684 CNN \u4e0d\u76f4\u63a5\u751f\u6210\u6700\u7ec8 RGB\uff0c\u800c\u662f\u9884\u6d4b\u7a7a\u95f4\u6ee4\u6ce2\u6838\u7684\u5f62\u72b6\u3001\u65b9\u5411\u548c\u65f6\u57df\u878d\u5408\u95e8\u63a7\u3002GPU \u9488\u5bf9\u6bcf\u4e2a\u9ad8\u5206\u8fa8\u7387\u8f93\u51fa\u50cf\u7d20\uff0c\u4ece\u4f4e\u5206\u8fa8\u7387\u5f53\u524d\u5e27\u7684\u90bb\u57df\u4e2d\u8fdb\u884c\u52a8\u6001\u9ad8\u65af\u91cd\u5efa\uff0c\u518d\u4e0e\u8fd0\u52a8\u8865\u507f\u540e\u7684\u9ad8\u5206\u8fa8\u7387\u5386\u53f2\u878d\u5408\u3002\u5386\u53f2\u7ed3\u679c\u901a\u8fc7\u9012\u5f52\u66f4\u65b0\u9010\u5e27\u79ef\u7d2f\u4e0d\u540c Jitter \u4f4d\u7f6e\u63d0\u4f9b\u7684\u4e9a\u50cf\u7d20\u4fe1\u606f\uff0c\u6700\u7ec8\u751f\u6210\u7a33\u5b9a\u7684\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u3002<\/p>\n\n\n

              “`<\/p>","protected":false},"excerpt":{"rendered":"

              1. DLSS \u603b\u89c8 \u7531\u4e8eNVIDIA\u7684\u5ba3\u4f20\u95ee\u9898\uff0cDLSS\u5b9e\u9645\u5e76\u4e0d\u4ec5\u4ec5\u662f\u72ed\u4e49SS\uff08super sampling […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[31,32],"class_list":["post-2263","post","type-post","status-publish","format-standard","hentry","category-8","tag-ai","tag-32"],"_links":{"self":[{"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/posts\/2263","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/comments?post=2263"}],"version-history":[{"count":16,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/posts\/2263\/revisions"}],"predecessor-version":[{"id":2285,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/posts\/2263\/revisions\/2285"}],"wp:attachment":[{"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/media?parent=2263"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/categories?post=2263"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/tags?post=2263"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}