來源：OpenCV學堂

OpenVINO2025 C#支持

開源的支持項目來自顏國進老師貢獻，已經(jīng)被OpenVINO官方收錄，項目網(wǎng)址：

https://gitee.com/linbei_cht/OpenVINO-CSharp-API

安裝非常容易，只要在VS2022里面點擊一下即可安裝。最新版本已經(jīng)是OpenVINO2025支持。

YOLO11實例分割

YOLO11是YOLOv5跟YOLOv8作者推出最新升級版本模型，支持分類、檢測、分割、姿態(tài)評估、OBB。這里以YOLO11實例分割模型為例，演示OpenVINO C#如何運行，YOLO11-seg模型的輸入與輸出。

代碼是我在OpenVINO-CSharp-API作者開源的YOLOv8對象檢測的代碼基礎(chǔ)上修改而成。調(diào)用檢測代碼如下：

publicvoidDetect()
{
 // Set the video path and model path
 stringvideo_path="D:/images/video/play_scoers.mp4";
 stringmodel_path="D:/python/yolov5-7.0/yolo11n-seg.onnx";
 // Create a new Core object and read the model
 Corecore=newCore();
 Modelmodel=core.read_model(model_path);
 CompiledModelcompiled_model=core.compile_model(model,"GPU");
 // Create a list of InferRequest objects
  Listrequests =newList{ compiled_model.create_infer_request(), compiled_model.create_infer_request() };
 // Create a new VideoCapture object and read the video
 VideoCapturecapture=newVideoCapture(video_path);
 if(!capture.IsOpened())
  {
    Console.WriteLine("Error: Video not found!");
   return;
  }
 Matframe=newMat();
 Matnext_frame=newMat();
  capture.Read(frame);
 floatscale=0.0f;
 float[] input_data = preprocess(frame, out scale);
  requests[0].get_input_tensor().set_data(input_data);
  requests[0].start_async();
 Stopwatchsw=newStopwatch();
 float[] total_infs =newfloat[3];
  ListclassList = File.ReadAllLines("D:/python/yolov5-7.0/classes.txt").Select(line => line.Trim()).ToList();
 while(true)
  {
   if(!capture.Read(next_frame))
    {
     break;
    }
    sw.Restart();
    input_data = preprocess(frame, out scale);
    requests[1].get_input_tensor().set_data(input_data);
    requests[1].start_async();
    requests[0].wait();
   float[] det_data = requests[0].get_tensor("output0").get_data(8400*116);
   float[] seg_data = requests[0].get_tensor("output1").get_data(32*160*160);
   Matrgb_mask=newMat(frame.Size(), frame.Type());
   DetResultresult=postprocess(det_data, seg_data, scale, rgb_mask);
    sw.Stop();
    total_infs[0] = sw.ElapsedMilliseconds;
    Cv2.PutText(frame,"Inference: "+ (1000.0/ total_infs[0]).ToString("0.00") +"FPS "+ (total_infs[0]).ToString("0.00") +"ms",newOpenCvSharp.Point(20,40), HersheyFonts.HersheyPlain,2,newScalar(255,0,255),2);
    result.update_lable(classList);
    Visualize.draw_det_result(result, frame);
    Cv2.AddWeighted(frame,0.5, rgb_mask,0.5,0, frame);
    Cv2.ImShow("C# YOLO11-OpenVINO-Seg演示 - OpenCV學堂", frame);
   // Press 'ESC' to exit the program
   if(Cv2.WaitKey(1) ==27)
    {
     break;
    }
    swap(requests);
    frame = next_frame;
    rgb_mask.Release();
  }
}

后處理實現(xiàn)細節(jié)

這個實現(xiàn)最大的坑在后處理部分，要基于全局編碼信息乘以每個檢測BOX區(qū)域的編碼信息，才可以解碼得到每個BOX對象的掩膜。實現(xiàn)的代碼如下：

Matroi_mask=roi_masks[index];
Matm=roi_mask * mask_info;
for(intcol=0; col < m.Cols; col++)
{
? ? m.At(0, col) = sigmoid_function(m.At(0, col));
}

最后根據(jù)得到掩膜直接設(shè)置BOX區(qū)域的顏色即可，代碼如下：

rgb_mask[box].SetTo(new Scalar(0,0,255), box_m);
re.add(classIds[index], confidences[index], positionBoxes[index]);

然后把得到RGB彩色掩膜圖像跟BOX框繪制圖像相加記得到最終輸出結(jié)果圖像。