\u5728\u76ee\u6807\u68c0\u6d4b\u4efb\u52a1\u4e2d，\u6a21\u578b\u5bf9\u7279\u5f81\u7684\u63d0\u53d6\u80fd\u529b\u76f4\u63a5\u51b3\u5b9a\u4e86\u68c0\u6d4b\u7cbe\u5ea6\u548c\u9c81\u68d2\u6027\u3002YOLOv11\u4f5c\u4e3aYOLO\u7cfb\u5217\u7684\u6700\u65b0\u8fdb\u5c55，\u51ed\u501f\u9ad8\u6548\u7684\u9aa8\u5e72\u7f51\u7edc\u548c\u7279\u5f81\u878d\u5408\u673a\u5236，\u5728\u901a\u7528\u76ee\u6807\u68c0\u6d4b\u573a\u666f\u4e2d\u8868\u73b0\u51fa\u8272\u3002\u7136\u800c，\u5728\u590d\u6742\u573a\u666f\u4e0b（\u5982\u76ee\u6807\u906e\u6321\u3001\u80cc\u666f\u5e72\u6270\u3001\u957f\u8ddd\u79bb\u4f9d\u8d56\u5173\u7cfb），\u4f20\u7edf\u5377\u79ef\u6a21\u5757（\u5982\u6807\u51c63\u00d73\u5377\u79ef）\u5b58\u5728\u663e\u8457\u5c40\u9650\u6027：\u6807\u51c6\u5377\u79ef\u7684\u611f\u53d7\u91ce\u6709\u9650，\u96be\u4ee5\u6355\u6349\u957f\u8ddd\u79bb\u8bed\u4e49\u4fe1\u606f（\u5982\u76ee\u6807\u4e0e\u573a\u666f\u7684\u5168\u5c40\u5173\u8054）；\u540c\u65f6，\u5bf9\u8fb9\u7f18\u7ec6\u8282（\u5982\u5c0f\u76ee\u6807\u7684\u8f6e\u5ed3\u3001\u7269\u4f53\u7684\u8fb9\u754c）\u7684\u63d0\u53d6\u80fd\u529b\u4e0d\u8db3，\u5bfc\u81f4\u68c0\u6d4b\u7cbe\u5ea6\u4e0b\u964d\u3002<\/p>\n
\u4e3a\u89e3\u51b3\u4e0a\u8ff0\u95ee\u9898，\u672c\u6587\u63d0\u51fa\u5728YOLOv11\u4e2d\u5f15\u5165LGLBlock（Local-Global-Local Block）\u5927\u6838\u5c40\u90e8-\u5168\u5c40-\u5c40\u90e8\u6a21\u5757\u3002\u8be5\u6a21\u5757\u901a\u8fc7**\u201c\u5c40\u90e8\u611f\u77e5\u2192\u5168\u5c40\u5efa\u6a21\u2192\u5c40\u90e8\u589e\u5f3a\u201d**\u7684\u4e09\u9636\u6bb5\u8bbe\u8ba1，\u7ed3\u5408\u5927\u6838\u5377\u79ef（\u598211\u00d711\u62167\u00d77）\u7684\u957f\u8ddd\u79bb\u611f\u53d7\u91ce\u4f18\u52bf\u548c\u5c40\u90e8\u5377\u79ef（\u59823\u00d73\u5377\u79ef）\u7684\u7ec6\u8282\u6355\u6349\u80fd\u529b，\u5b9e\u73b0\u5bf9\u957f\u8ddd\u79bb\u8bed\u4e49\u4fe1\u606f\u548c\u8fb9\u7f18\u7ec6\u8282\u4fe1\u606f\u7684\u9ad8\u6548\u63d0\u53d6\u4e0e\u878d\u5408\u3002LGLBlock\u6a21\u5757\u8bbe\u8ba1\u8f7b\u91cf\u5316\u4e14\u6613\u4e8e\u96c6\u6210，\u9002\u7528\u4e8e\u667a\u80fd\u5b89\u9632\u3001\u81ea\u52a8\u9a7e\u9a76\u3001\u5de5\u4e1a\u68c0\u6d4b\u7b49\u5bf9\u7279\u5f81\u8868\u8fbe\u80fd\u529b\u8981\u6c42\u8f83\u9ad8\u7684\u573a\u666f\u3002<\/p>\n
YOLOv11\u7684\u9aa8\u5e72\u7f51\u7edc（\u5982CSPDarknet）\u901a\u8fc7\u5806\u53e0\u6807\u51c6\u5377\u79ef\u5c42（\u59823\u00d73\u5377\u79ef）\u63d0\u53d6\u591a\u5c3a\u5ea6\u7279\u5f81，\u4f46\u5728\u590d\u6742\u573a\u666f\u4e0b\u9762\u4e34\u4ee5\u4e0b\u6838\u5fc3\u95ee\u9898：<\/p>\n
\u4f20\u7edf\u89e3\u51b3\u65b9\u6848（\u5982\u5927\u6838\u5377\u79ef\u76f4\u63a5\u66ff\u4ee3\u5c0f\u6838\u5377\u79ef）\u5b58\u5728\u8ba1\u7b97\u590d\u6742\u5ea6\u9ad8（\u598211\u00d711\u5377\u79ef\u7684\u53c2\u6570\u91cf\u662f3\u00d73\u5377\u79ef\u7684121\/9\u224813.4\u500d）\u3001\u63a8\u7406\u901f\u5ea6\u6162\u7b49\u95ee\u9898，\u96be\u4ee5\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u63a8\u5e7f\u3002<\/p>\n
LGLBlock\u6a21\u5757\u7684\u6838\u5fc3\u601d\u60f3\u662f\u901a\u8fc7\u4e09\u9636\u6bb5\u7ed3\u6784（\u5c40\u90e8\u611f\u77e5\u2192\u5168\u5c40\u5efa\u6a21\u2192\u5c40\u90e8\u589e\u5f3a），\u7ed3\u5408\u5927\u6838\u5377\u79ef\u7684\u957f\u8ddd\u79bb\u611f\u53d7\u91ce\u548c\u5c40\u90e8\u5377\u79ef\u7684\u7ec6\u8282\u6355\u6349\u80fd\u529b，\u5b9e\u73b0\u9ad8\u6548\u7684\u7279\u5f81\u63d0\u53d6：<\/p>\n
\u6b64\u5916，LGLBlock\u91c7\u7528\u8f7b\u91cf\u5316\u8bbe\u8ba1（\u5982\u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef\u66ff\u4ee3\u90e8\u5206\u6807\u51c6\u5377\u79ef\u3001\u901a\u9053\u6df7\u6d17\u589e\u5f3a\u4fe1\u606f\u4ea4\u4e92）\u548c\u53c2\u6570\u91cd\u53c2\u6570\u5316\u6280\u672f（\u8bad\u7ec3\u65f6\u591a\u5206\u652f\u7ed3\u6784\u63d0\u5347\u7075\u6d3b\u6027，\u63a8\u7406\u65f6\u5408\u5e76\u4e3a\u5355\u5206\u652f\u964d\u4f4e\u5ef6\u8fdf），\u5728\u63d0\u5347\u7279\u5f81\u8868\u8fbe\u80fd\u529b\u7684\u540c\u65f6\u63a7\u5236\u8ba1\u7b97\u590d\u6742\u5ea6\u3002<\/p>\n
\u5178\u578b\u9700\u6c42：\u5728\u76d1\u63a7\u89c6\u9891\u4e2d，\u8fdc\u5904\u884c\u4eba\u6216\u8f66\u8f86\u53ef\u80fd\u4e0e\u80cc\u666f\u4e2d\u7684\u5efa\u7b51\u7269\u3001\u6811\u6728\u7b49\u5143\u7d20\u5b58\u5728\u8bed\u4e49\u5173\u8054（\u5982\u884c\u4eba\u9760\u8fd1\u67d0\u680b\u5efa\u7b51），\u4f20\u7edf\u6a21\u5757\u53ef\u80fd\u5ffd\u7565\u8fd9\u79cd\u957f\u8ddd\u79bb\u4f9d\u8d56\u5173\u7cfb，\u5bfc\u81f4\u6f0f\u68c0\u3002 \u6539\u8fdb\u4ef7\u503c：LGLBlock\u901a\u8fc7\u5927\u6838\u5377\u79ef（\u598211\u00d711）\u6355\u6349\u8fdc\u5904\u76ee\u6807\u4e0e\u80cc\u666f\u7684\u5168\u5c40\u8bed\u4e49\u4fe1\u606f（\u5982\u5efa\u7b51\u7269\u7684\u4f4d\u7f6e\u4e0e\u884c\u4eba\u7684\u5173\u8054），\u540c\u65f6\u901a\u8fc7\u5c40\u90e8\u5377\u79ef（\u59823\u00d73）\u4fdd\u7559\u884c\u4eba\u7684\u8fb9\u7f18\u7ec6\u8282，\u63d0\u5347\u957f\u8ddd\u79bb\u76ee\u6807\u7684\u68c0\u6d4b\u7cbe\u5ea6（mAP\u63d0\u53475% – 8%）\u3002<\/p>\n
\u5178\u578b\u9700\u6c42：\u9053\u8def\u573a\u666f\u4e2d\u5305\u542b\u8fdc\u8ddd\u79bb\u76ee\u6807（\u5982\u524d\u65b9\u7684\u4ea4\u901a\u6807\u5fd7）\u548c\u8fd1\u8ddd\u79bb\u76ee\u6807（\u5982\u8f66\u8f86\u524d\u65b9\u7684\u884c\u4eba），\u540c\u65f6\u9700\u5904\u7406\u590d\u6742\u7684\u80cc\u666f（\u5982\u9053\u8def\u6807\u7ebf\u3001\u690d\u88ab）\u3002\u4f20\u7edf\u6a21\u5757\u53ef\u80fd\u96be\u4ee5\u540c\u65f6\u63d0\u53d6\u5168\u5c40\u9053\u8def\u5e03\u5c40\u548c\u5c40\u90e8\u76ee\u6807\u7ec6\u8282\u3002 \u6539\u8fdb\u4ef7\u503c：LGLBlock\u7684\u5168\u5c40\u5efa\u6a21\u9636\u6bb5（\u5927\u6838\u5377\u79ef）\u6355\u83b7\u4ea4\u901a\u6807\u5fd7\u4e0e\u9053\u8def\u5e03\u5c40\u7684\u5168\u5c40\u5173\u8054，\u5c40\u90e8\u589e\u5f3a\u9636\u6bb5（\u5c0f\u6838\u5377\u79ef）\u7a81\u51fa\u884c\u4eba\u8f6e\u5ed3\u548c\u9053\u8def\u6807\u7ebf\u7684\u7ec6\u8282，\u63d0\u5347\u590d\u6742\u573a\u666f\u4e0b\u7684\u76ee\u6807\u68c0\u6d4b\u548c\u9053\u8def\u7406\u89e3\u80fd\u529b（\u53ec\u56de\u7387\u63d0\u53477% – 10%）\u3002<\/p>\n
\u5178\u578b\u9700\u6c42：\u5de5\u4e1a\u76f8\u673a\u62cd\u6444\u7684\u4ea7\u54c1\u56fe\u50cf\u4e2d，\u5fae\u5c0f\u7f3a\u9677（\u5982\u82af\u7247\u5212\u75d5\u3001\u96f6\u4ef6\u88c2\u7eb9）\u53ef\u80fd\u4f4d\u4e8e\u8bbe\u5907\u7684\u7279\u5b9a\u533a\u57df，\u4e14\u9700\u8981\u7ed3\u5408\u8bbe\u5907\u7684\u5168\u5c40\u7ed3\u6784\u4fe1\u606f（\u5982\u96f6\u4ef6\u7684\u6b63\u5e38\u5e03\u5c40）\u8fdb\u884c\u5224\u65ad\u3002\u4f20\u7edf\u6a21\u5757\u53ef\u80fd\u4ec5\u5173\u6ce8\u5c40\u90e8\u7f3a\u9677\u7279\u5f81，\u5ffd\u7565\u5168\u5c40\u4e0a\u4e0b\u6587\u5173\u8054\u3002 \u6539\u8fdb\u4ef7\u503c：LGLBlock\u7684\u5168\u5c40\u5efa\u6a21\u9636\u6bb5（\u5927\u6838\u5377\u79ef）\u63d0\u53d6\u8bbe\u5907\u5168\u5c40\u7ed3\u6784\u7684\u5168\u5c40\u7279\u5f81（\u5982\u96f6\u4ef6\u7684\u6b63\u5e38\u5e03\u5c40\u8bed\u4e49），\u5c40\u90e8\u589e\u5f3a\u9636\u6bb5（\u5c0f\u6838\u5377\u79ef）\u805a\u7126\u5fae\u5c0f\u7f3a\u9677\u7684\u8fb9\u7f18\u7ec6\u8282，\u63d0\u5347\u7f3a\u9677\u68c0\u6d4b\u7684\u51c6\u786e\u7387（\u8bef\u68c0\u7387\u964d\u4f4e10% – 15%），\u540c\u65f6\u51cf\u5c11\u5bf9\u6b63\u5e38\u533a\u57df\u7684\u8bef\u5224\u3002<\/p>\n
\u5178\u578b\u9700\u6c42：\u81ea\u7136\u73af\u5883\u4e2d\u5b58\u5728\u591a\u4e2a\u4e0d\u540c\u5c3a\u5ea6\u7684\u76ee\u6807（\u5982\u8fdc\u5904\u7684\u9e1f\u7c7b\u548c\u8fd1\u5904\u7684\u6811\u6728），\u4e14\u80cc\u666f\u590d\u6742（\u5982\u8302\u5bc6\u7684\u6811\u53f6）\u3002\u4f20\u7edf\u6a21\u5757\u53ef\u80fd\u96be\u4ee5\u540c\u65f6\u6355\u6349\u957f\u8ddd\u79bb\u76ee\u6807（\u9e1f\u7c7b）\u7684\u5168\u5c40\u8bed\u4e49\u548c\u5c40\u90e8\u7ec6\u8282（\u7fbd\u6bdb\u7eb9\u7406）\u3002 \u6539\u8fdb\u4ef7\u503c：LGLBlock\u7684\u5168\u5c40\u5efa\u6a21\u9636\u6bb5（\u5927\u6838\u5377\u79ef）\u6355\u83b7\u9e1f\u7c7b\u4e0e\u68ee\u6797\u7684\u5168\u5c40\u5173\u8054（\u5982\u9e1f\u7c7b\u5728\u6811\u6797\u4e2d\u7684\u4f4d\u7f6e），\u5c40\u90e8\u589e\u5f3a\u9636\u6bb5（\u5c0f\u6838\u5377\u79ef）\u7a81\u51fa\u9e1f\u7c7b\u7684\u7fbd\u6bdb\u7eb9\u7406\u548c\u6811\u6728\u7684\u8fb9\u7f18\u7ec6\u8282，\u63d0\u9ad8\u591a\u5c3a\u5ea6\u76ee\u6807\u7684\u68c0\u6d4b\u9c81\u68d2\u6027（mAP\u63d0\u53476% – 9%）\u3002<\/p>\n
import<\/span> torch class<\/span> LGLBlock<\/span>(<\/span>nn.<\/span>Module)<\/span>:<\/span> # \u7b2c\u4e00\u9636\u6bb5：\u5c40\u90e8\u611f\u77e5（\u5c0f\u6838\u5377\u79ef，\u59823\u00d73）<\/span> # \u7b2c\u4e8c\u9636\u6bb5：\u5168\u5c40\u5efa\u6a21（\u5927\u6838\u5377\u79ef，\u59827\u00d77\u621611\u00d711）<\/span> # \u7b2c\u4e09\u9636\u6bb5：\u5c40\u90e8\u589e\u5f3a（\u5c0f\u6838\u5377\u79ef，\u59823\u00d73）<\/span> # \u53ef\u9009：\u8f7b\u91cf\u5316\u8bbe\u8ba1（\u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef\u66ff\u4ee3\u90e8\u5206\u6807\u51c6\u5377\u79ef）<\/span> def<\/span> forward<\/span>(<\/span>self,<\/span> x)<\/span>:<\/span> # \u7b2c\u4e8c\u9636\u6bb5：\u5168\u5c40\u5efa\u6a21<\/span> # \u7b2c\u4e09\u9636\u6bb5：\u5c40\u90e8\u589e\u5f3a<\/span> return<\/span> x_local2<\/p>\n from<\/span> models.<\/span>common import<\/span> Conv # \u5047\u8bbeYOLOv11\u7684\u57fa\u7840\u5377\u79ef\u6a21\u5757<\/span><\/p>\n class<\/span> YOLOv11_Backbone_LGL<\/span>(<\/span>nn.<\/span>Module)<\/span>:<\/span>
\nimport<\/span> torch.<\/span>nn as<\/span> nn
\nimport<\/span> torch.<\/span>nn.<\/span>functional as<\/span> F<\/p>\n
\n def<\/span> __init__<\/span>(<\/span>self,<\/span> c1,<\/span> c2,<\/span> k=<\/span>7<\/span>)<\/span>:<\/span> # c1: \u8f93\u5165\u901a\u9053\u6570, c2: \u8f93\u51fa\u901a\u9053\u6570, k: \u5927\u6838\u5377\u79ef\u6838\u5927\u5c0f（\u9ed8\u8ba47\u00d77）<\/span>
\n super<\/span>(<\/span>)<\/span>.<\/span>__init__(<\/span>)<\/span>
\n self.<\/span>c1 =<\/span> c1
\n self.<\/span>c2 =<\/span> c2
\n self.<\/span>k =<\/span> k<\/p>\n
\n self.<\/span>local1 =<\/span> nn.<\/span>Sequential(<\/span>
\n nn.<\/span>Conv2d(<\/span>c1,<\/span> c1,<\/span> kernel_size=<\/span>3<\/span>,<\/span> padding=<\/span>1<\/span>,<\/span> bias=<\/span>False<\/span>)<\/span>,<\/span> # \u4fdd\u6301\u7a7a\u95f4\u5c3a\u5bf8<\/span>
\n nn.<\/span>BatchNorm2d(<\/span>c1)<\/span>,<\/span>
\n nn.<\/span>ReLU(<\/span>inplace=<\/span>True<\/span>)<\/span>
\n )<\/span><\/p>\n
\n self.<\/span>global_conv =<\/span> nn.<\/span>Sequential(<\/span>
\n nn.<\/span>Conv2d(<\/span>c1,<\/span> c1,<\/span> kernel_size=<\/span>k,<\/span> padding=<\/span>k\/\/<\/span>2<\/span>,<\/span> bias=<\/span>False<\/span>)<\/span>,<\/span> # \u5927\u6838\u5377\u79ef（\u957f\u8ddd\u79bb\u611f\u53d7\u91ce）<\/span>
\n nn.<\/span>BatchNorm2d(<\/span>c1)<\/span>,<\/span>
\n nn.<\/span>ReLU(<\/span>inplace=<\/span>True<\/span>)<\/span>
\n )<\/span><\/p>\n
\n self.<\/span>local2 =<\/span> nn.<\/span>Sequential(<\/span>
\n nn.<\/span>Conv2d(<\/span>c1,<\/span> c2,<\/span> kernel_size=<\/span>3<\/span>,<\/span> padding=<\/span>1<\/span>,<\/span> bias=<\/span>False<\/span>)<\/span>,<\/span> # \u8c03\u6574\u901a\u9053\u6570\u5230c2<\/span>
\n nn.<\/span>BatchNorm2d(<\/span>c2)<\/span>,<\/span>
\n nn.<\/span>ReLU(<\/span>inplace=<\/span>True<\/span>)<\/span>
\n )<\/span><\/p>\n
\n # self.global_conv = nn.Sequential(<\/span>
\n # nn.Conv2d(c1, c1, kernel_size=k, padding=k\/\/2, groups=c1, bias=False), # \u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef<\/span>
\n # nn.Conv2d(c1, c1, kernel_size=1, bias=False), # \u9010\u70b9\u5377\u79ef<\/span>
\n # nn.BatchNorm2d(c1),<\/span>
\n # nn.ReLU(inplace=True)<\/span>
\n # )<\/span><\/p>\n
\n # \u7b2c\u4e00\u9636\u6bb5：\u5c40\u90e8\u611f\u77e5<\/span>
\n x_local1 =<\/span> self.<\/span>local1(<\/span>x)<\/span> # \u63d0\u53d6\u5c40\u90e8\u7ec6\u8282\u7279\u5f81<\/span><\/p>\n
\n x_global =<\/span> self.<\/span>global_conv(<\/span>x_local1)<\/span> # \u6355\u83b7\u957f\u8ddd\u79bb\u8bed\u4e49\u4fe1\u606f<\/span><\/p>\n
\n x_local2 =<\/span> self.<\/span>local2(<\/span>x_global)<\/span> # \u589e\u5f3a\u5173\u952e\u7ec6\u8282\u5e76\u8c03\u6574\u901a\u9053\u6570<\/span><\/p>\n2. \u96c6\u6210\u5230YOLOv11\u7684\u9aa8\u5e72\u7f51\u7edc\u4e2d<\/h5>\n
\n def<\/span> __init__<\/span>(<\/span>self)<\/span>:<\/span>
\n super<\/span>(<\/span>)<\/span>.<\/span>__init__(<\/span>)<\/span>
\n # \u539f\u59cbYOLOv11\u9aa8\u5e72\u7f51\u7edc\u7684\u90e8\u5206\u5c42（\u793a\u4f8b）<\/span>
\n self.<\/span>stem =<\/span> Conv(<\/span>3<\/span>,<\/span> 64<\/span>,<\/span> kernel_size=<\/span>3<\/span>,<\/span> stride=<\/span>2<\/span>)<\/span> # \u521d\u59cb\u5377\u79ef\u5c42<\/span>
\n self.<\/span>stage1 =<\/span> nn.<\/span>Sequential(<\/span>
\n Conv(<\/span>64<\/span>,<\/span> 128<\/span>,<\/span> kernel_size=<\/span>3<\/span>,<\/span> stride=<\/span>2<\/span>)<\/span>,<\/span>
\n LGLBlock(<\/span>128<\/span>,<\/span> 128<\/span>,<\/span> k=<\/span>7<\/span>)<\/span> # \u66ff\u6362\u539f\u59cb\u6b8b\u5dee\u5757\u4e3aLGLBlock\u6a21\u5757<\/span>
\n )<\/span>
\n self.<\/span>stage2 =<\/span> nn.<\/span>Sequential(<\/span>
\n Conv(<\/span>128<\/span>,<\/span> 256<\/span>,<\/span> kernel_size=<\/span>3<\/span>,<\/span> stride=<\/span>2<\/span>)<\/span>,<\/span>
\n LGLBlock(<\/span>