1

.gitignore

@@ -1,4 +1,6 @@
 /logs
 /temp
 /.vs
-zhipuzi_pos_windows
+tmp
+db
+log

bin/Win32/Debug/zhipuzi_pos_windows/ai/cls.names

@@ -0,0 +1,2 @@
+car
+bus

bin/Win32/Release/zhipuzi_pos_windows/ai/cls.names

@@ -0,0 +1,2 @@
+car
+bus

zhipuzi_pos_windows/ai/YoloFeatureExtractor.cpp

@@ -11,7 +11,7 @@
 #include "../tool/debuglog.h"
 
 YoloFeatureExtractor::YoloFeatureExtractor(const std::string& modelPath, const std::string& classesPath)
-	: inputWidth(448), inputHeight(448)
+	: inputWidth(800), inputHeight(800)
 {
 	net = cv::dnn::readNetFromONNX(modelPath);
 	loadClassNames(classesPath);
@@ -85,8 +85,15 @@ std::vector<float> YoloFeatureExtractor::extractFeatures(const std::string& imag
 
 		cv::resize(image, image, cv::Size(inputWidth, inputHeight));
 
+		// 2. 转换为blob（归一化+通道转换）
+		//均值和标准差使用ImageNet标准值，与YOLO预训练分类模型一致
+		cv::Scalar mean = cv::Scalar(0.485 * 255, 0.456 * 255, 0.406 * 255);
+		cv::Scalar std = cv::Scalar(0.229 * 255, 0.224 * 255, 0.225 * 255);
+
 		cv::Mat blob;
-		cv::dnn::blobFromImage(image, blob, 1.0 / 255.0, cv::Size(inputWidth, inputHeight), cv::Scalar(0, 0, 0), true, false);
+		cv::dnn::blobFromImage(image, blob, 1.0, cv::Size(inputWidth, inputHeight), mean, true, false);
+		blob /= std; // 应用标准差归一化
+
 		net.setInput(blob);
 
 		DEBUG_LOG::debug_printf("blob 形状：%d×%d×%d×%d\n", blob.size[0], blob.size[1], blob.size[2], blob.size[3]);
@@ -94,44 +101,61 @@ std::vector<float> YoloFeatureExtractor::extractFeatures(const std::string& imag
 		auto time_2 = std::chrono::high_resolution_clock::now();
 
 		std::vector<cv::String> layerNames = net.getLayerNames();
-		std::vector<cv::String> outputNames;
 
 		std::vector<float> features;
 
-		
-		// 选择GAP层（对于yolo2026，通常是倒数第6层）的输出作为特征向量
-		outputNames.push_back(layerNames[layerNames.size() - 6]);
-
-		std::vector<cv::Mat> outputs;
-		net.forward(outputs, outputNames);
-
+		cv::Mat featureMat = net.forward("onnx_node!/model.10/pool/GlobalAveragePool");
 		// 检查输出是否有效
-		if (outputs.empty() || outputs[0].empty())
+		if (featureMat.empty())
 		{
 			throw std::runtime_error("模型前向传播未产生有效输出");
 		}
 
 		// 获取GAP层输出并转换为特征向量
-		cv::Mat featuresMat = outputs[0];
-		DEBUG_LOG::debug_printf("原始特征形状：%d×%d，类型：%d（CV_32F=5）\n", featuresMat.cols, featuresMat.rows, featuresMat.type());
+		//DEBUG_LOG::debug_printf("原始特征形状：%d×%d，类型：%d（CV_32F=5）\n", featureMat.cols, featureMat.rows, featureMat.type());
 			
-		cv::Mat featuresMatVec = featuresMat.reshape(1, 1);
-		DEBUG_LOG::debug_printf("重塑后特征形状：%d×%d，类型：%d\n", featuresMatVec.cols, featuresMatVec.rows, featuresMatVec.type());
+		featureMat = featureMat.reshape(1, 1);
+		//DEBUG_LOG::debug_printf("重塑后特征形状：%d×%d，类型：%d\n", featureMat.cols, featureMat.rows, featureMat.type());
 
-		float norm_before = cv::norm(featuresMatVec, cv::NORM_L2);
-		DEBUG_LOG::debug_printf("归一化前 norm：%.6f\n", norm_before);
+		//float norm_before = cv::norm(featureMat, cv::NORM_L2);
+		//DEBUG_LOG::debug_printf("归一化前 norm：%.6f\n", norm_before);
 
-		//normalizeL2(featuresMatVec);
-		cv::normalize(featuresMat, featuresMatVec, 1.0, 0.0, cv::NORM_L2);
+		//normalizeL2(featureMat);
+		//cv::normalize(featureMat, featureMat, 1.0, 0.0, cv::NORM_L2);
 
-		float norm_after = cv::norm(featuresMatVec, cv::NORM_L2);
-		DEBUG_LOG::debug_printf("归一化后 norm：%.6f\n", norm_after);
+		//float norm_after = cv::norm(featureMat, cv::NORM_L2);
+		//DEBUG_LOG::debug_printf("归一化后 norm：%.6f\n", norm_after);
 
-		features.reserve(features.size() + featuresMatVec.total());
-		for (size_t j = 0; j < featuresMatVec.total(); ++j)
+		// ===================== 5. 特征向量处理 =====================
+		// 将Mat格式的特征转换为vector<float>（方便后续计算/存储）
+		std::vector<float> feature_vector;
+		feature_vector.assign((float *)featureMat.data, (float *)featureMat.data + featureMat.total());
+
+		// -------------------- 新增：L2归一化 -------------------
+		// 1. 计算L2范数
+		/*
+		float l2_norm = 0.0f;
+		for (float val : feature_vector)
 		{
-			features.push_back(featuresMatVec.at<float>(j));
+			l2_norm += val * val;
 		}
+		l2_norm = sqrt(l2_norm);
+
+		// 2. 避免除以0（极端情况，特征全为0）
+		if (fabs(l2_norm) < 1e-8)
+		{
+			l2_norm = 1e-8;
+		}
+
+		// 3. 归一化：每个元素除以L2范数
+		std::vector<float> normalized_feature;
+		for (float val : feature_vector)
+		{
+			normalized_feature.push_back(val / l2_norm);
+		}
+		// -------------------- L2归一化结束 --------------------
+		*/
+
 
 		// 转换为std::vector<float>
 		//features = std::vector<float>(featuresMatVec.begin<float>(), featuresMatVec.end<float>());
@@ -150,7 +174,7 @@ std::vector<float> YoloFeatureExtractor::extractFeatures(const std::string& imag
 		std::wstring msg4 = L"总耗时: " + std::to_wstring(totalDuration.count()) + L" 毫秒";
 		DEBUG_LOG(msg4.c_str());
 
-		return features;
+		return feature_vector;
 	}
 	catch (const std::exception& e)
 	{

zhipuzi_pos_windows/ai/test.cpp

@@ -25,7 +25,7 @@ int AITest()
 		//用于测试的图片目录
 		std::string galleryDir = (mainDir.parent_path().parent_path().parent_path().parent_path() /"res"/"images").string();       // 图库目录路径
 
-		std::string modelPath = sMainDir + "/ai/best_448.onnx";           // YOLO2026模型路径
+		std::string modelPath = sMainDir + "/ai/best_800.onnx";           // YOLO2026模型路径
 		std::string classesPath = sMainDir + "/ai/cls.names";             // 类别文件路径
 		std::string searchImagePath = sMainDir + "/3.jpg"; // 搜索图片路径