Runtime跑YOLO

这是我们产线验证过的核心代码，包括模型加载、图像预处理、推理、结果解析，拿来就能用。

1.

安装NuGet包

在NuGet包管理器里安装：

• Microsoft.ML.OnnxRuntime（推理引擎）
• Microsoft.ML.OnnxRuntime.Managed（托管API）
• Emgu.CV、Emgu.CV.runtime.windows（图像处理）

2.

核心检测类代码

usingMicrosoft.ML.OnnxRuntime;usingMicrosoft.ML.OnnxRuntime.Tensors;usingEmgu.CV;usingEmgu.CV.Structure;usingSystem.Collections.Concurrent;namespaceLabelDetection.Services;publicclassYoloDetector{privatereadonlyInferenceSession_session;privatereadonlystring[]_classNames={"正常标签","错贴标签","无标签"};privateconstintInputSize=640;//
模型输入尺寸，和导出时一致privateconstfloatConfidenceThreshold=0.5f;//
置信度阈值privateconstfloatIouThreshold=0.45f;//
NMS的IOU阈值publicYoloDetector(stringmodelPath){//
配置Session，开启图优化varsessionOptions=newSessionOptions();sessionOptions.GraphOptimizationLevel=GraphOptimizationLevel.ORT_ENABLE_ALL;//
如果有GPU，取消下面这行的注释//
sessionOptions.AppendExecutionProvider_CUDA(0);_session=newInferenceSession(modelPath,sessionOptions);}///
<summary>///
图像预处理：调整尺寸、归一化、转成Tensor///
</summary>privateDenseTensor<float>PreprocessImage(Matimage){//
调整图像尺寸到InputSize
InputSizeMatresized=newMat();CvInvoke.Resize(image,resized,newSize(InputSize,InputSize));//
转成RGB（Emgu
CV默认是BGR）Matrgb=newMat();CvInvoke.CvtColor(resized,rgb,ColorConversion.Bgr2Rgb);//
归一化（0-255
0-1）rgb.ConvertTo(rgb,DepthType.Cv32F,1.0/255.0);//
channels,
width]vartensor=newDenseTensor<float>(new[]{1,3,InputSize,InputSize});vardata=rgb.GetData();for(inty=0;y<InputSize;y++){for(intx=0;x<InputSize;x++){tensor[0,0,y,x]=data[y,x,0];//
Rtensor[0,1,y,x]=data[y,x,1];//
Gtensor[0,2,y,x]=data[y,x,2];//
B}}returntensor;}///
<summary>///
执行检测///
</summary>publicList<DetectionResult>Detect(Matimage){//
预处理varinputTensor=PreprocessImage(image);//
构建输入varinputs=newList<NamedOnnxValue>{NamedOnnxValue.CreateFromTensor("images",inputTensor)};//
执行推理usingvarresults=_session.Run(inputs);//
解析输出varoutputTensor=results.First().AsEnumerable<float>().ToArray();vardetections=ParseOutput(outputTensor,image.Width,image.Height);//
NMS去除重复框detections=NMS(detections);returndetections;}///
<summary>///
</summary>privateList<DetectionResult>ParseOutput(float[]output,intoriginalWidth,intoriginalHeight){vardetections=newList<DetectionResult>();intnumDetections=output.Length/(4+1+_classNames.Length);//
4个坐标+1个置信度+类别数for(inti=0;i<numDetections;i++){intoffset=i*(4+1+_classNames.Length);//
置信度floatconfidence=output[offset+4];if(confidence<ConfidenceThreshold)continue;//
类别intclassId=0;floatmaxClassScore=0;for(intj=0;j<_classNames.Length;j++){floatscore=output[offset+5+j];if(score>maxClassScore){maxClassScore=score;classId=j;}}//
坐标：x_center,
y2floatxCenter=output[offset]*originalWidth/InputSize;floatyCenter=output[offset+1]*originalHeight/InputSize;floatwidth=output[offset+2]*originalWidth/InputSize;floatheight=output[offset+3]*originalHeight/InputSize;floatx1=xCenter-width/2;floaty1=yCenter-height/2;floatx2=xCenter+width/2;floaty2=yCenter+height/2;detections.Add(newDetectionResult{ClassId=classId,ClassName=_classNames[classId],Confidence=confidence,X1=x1,Y1=y1,X2=x2,Y2=y2});}returndetections;}///
<summary>///
</summary>privateList<DetectionResult>NMS(List<DetectionResult>detections){varsorted=detections.OrderByDescending(d=>d.Confidence).ToList();varkept=newList<DetectionResult>();while(sorted.Count>0){varcurrent=sorted[0];kept.Add(current);sorted.RemoveAt(0);for(inti=sorted.Count-1;i>=0;i--){floatiou=CalculateIOU(current,sorted[i]);if(iou>IouThreshold){sorted.RemoveAt(i);}}}returnkept;}///
<summary>///
</summary>privatefloatCalculateIOU(DetectionResulta,DetectionResultb){floatx1=Math.Max(a.X1,b.X1);floaty1=Math.Max(a.Y1,b.Y1);floatx2=Math.Min(a.X2,b.X2);floaty2=Math.Min(a.Y2,b.Y2);floatintersection=Math.Max(0,x2-x1)*Math.Max(0,y2-y1);floatareaA=(a.X2-a.X1)*(a.Y2-a.Y1);floatareaB=(b.X2-b.X1)*(b.Y2-b.Y1);floatunion=areaA+areaB-intersection;returnintersection/union;}}///
<summary>///
</summary>publicclassDetectionResult{publicintClassId{get;set;}publicstringClassName{get;set;}publicfloatConfidence{get;set;}publicfloatX1{get;set;}publicfloatY1{get;set;}publicfloatX2{get;set;}publicfloatY2{get;set;}}