【OpenVINO™】在 C# 中使用OpenVINO™ 部署PP-YOLOE实现物体检测

我冷笑西风

 前言OpenVINO™ C# API 是一个 OpenVINO™ 的 .Net wrapper，应用最新的 OpenVINO™ 库开发，通过 OpenVINO™ C API 实现 .Net 对 OpenVINO™ Runtime 调用，使用习惯与 OpenVINO™ C++ API 一致。OpenVINO™ C# API 由于是基于 OpenVINO™ 开发，所支持的平台与 OpenVINO™ 完全一致，具体信息可以参考 OpenVINO™。通过使用 OpenVINO™ C# API，可以在 .NET、.NET Framework等框架下使用 C# 语言实现深度学习模型在指定平台推理加速。
OpenVINO™ C# API 项目链接为：

1. https://github.com/guojin-yan/OpenVINO-CSharp-API.git

复制代码

项目源码链接为：

1. https://github.com/guojin-yan/OpenVINO-CSharp-API-Samples.git

复制代码

1. 简介

    PP-YOLOE是基于PP-YOLOv2的优秀单级无锚模型，超越了各种流行的YOLO模型。PP-YOLOE有一系列型号，命名为s/m/l/x，通过宽度乘数和深度乘数进行配置。PP-YOLOE避免使用特殊的运算符，如可变形卷积或矩阵NMS，以便友好地部署在各种硬件上。 在本文中，我们将使用OpenVINO™ C# API 部署 PP-YOLOE实现物体检测。
2. 项目环境与依赖

     该项目中所需依赖已经支持通过NuGet Package进行安装，在该项目中，需要安装以下NuGet Package：

• OpenVINO C# API NuGet Package：
  

1. OpenVINO.CSharp.API
2. OpenVINO.runtime.win
3. OpenVINO.CSharp.API.Extensions
4. OpenVINO.CSharp.API.Extensions.OpenCvSharp

复制代码

• OpenCvSharp NuGet Package：
  

1. OpenCvSharp4
2. OpenCvSharp4.Extensions
3. OpenCvSharp4.runtime.win

复制代码

3. 项目输出

     项目使用的是控制台输出，运行后输出如下所示：

1. <00:00:00> Sending http request to https://github.com/guojin-yan/OpenVINO-CSharp-API-Samples/releases/download/Model/ppyoloe_plus_crn_l_80e_coco.tar.
2. <00:00:02> Http Response Accquired.
3. <00:00:02> Total download length is 199.68 Mb.
4. <00:00:02> Download Started.
5. <00:00:02> File created.
6. <00:02:03> Downloading: [■■■■■■■■■■] 100% <00:02:03 1.81 Mb/s> 199.68 Mb/199.68 Mb downloaded.
7. <00:02:03> File Downloaded, saved in E:\GitSpace\OpenVINO-CSharp-API-Samples\model_samples\ppyoloe\ppyoloe_opencvsharp\bin\Release\net6.0\model\ppyoloe_plus_crn_l_80e_coco.tar.
8. <00:00:00> Sending http request to https://github.com/guojin-yan/OpenVINO-CSharp-API-Samples/releases/download/Image/test_det_02.jpg.
9. <00:00:02> Http Response Accquired.
10. <00:00:02> Total download length is 0.16 Mb.
11. <00:00:02> Download Started.
12. <00:00:02> File created.
13. <00:00:02> Downloading: [■■■■■■■■■■] 100% <00:00:02 0.06 Mb/s> 0.16 Mb/0.16 Mb downloaded.
14. <00:00:02> File Downloaded, saved in E:\GitSpace\OpenVINO-CSharp-API-Samples\model_samples\ppyoloe\ppyoloe_opencvsharp\bin\Release\net6.0\model\test_image.jpg.
15. [ INFO ] Inference device: CPU
16. [ INFO ] Start RT-DETR model inference.
17. [ INFO ] 1. Initialize OpenVINO Runtime Core success, time spend: 4.5204ms.
18. [ INFO ] 2. Read inference model success, time spend: 228.4451ms.
19. [ INFO ] Inference Model
20. [ INFO ]   Model name: Model0
21. [ INFO ]   Input:
22. [ INFO ]      name: scale_factor
23. [ INFO ]      type: float
24. [ INFO ]      shape: Shape : {?,2}
25. [ INFO ]      name: image
26. [ INFO ]      type: float
27. [ INFO ]      shape: Shape : {?,3,640,640}
28. [ INFO ]   Output:
29. [ INFO ]      name: multiclass_nms3_0.tmp_0
30. [ INFO ]      type: float
31. [ INFO ]      shape: Shape : {?,6}
32. [ INFO ]      name: multiclass_nms3_0.tmp_2
33. [ INFO ]      type: int32_t
34. [ INFO ]      shape: Shape : {?}
35. [ INFO ] 3. Loading a model to the device success, time spend:501.0716ms.
36. [ INFO ] 4. Create an infer request success, time spend:0.2663ms.
37. [ INFO ] 5. Process input images success, time spend:30.1001ms.
38. [ INFO ] 6. Set up input data success, time spend:2.3631ms.
39. [ INFO ] 7. Do inference synchronously success, time spend:286.1085ms.
40. [ INFO ] 8. Get infer result data success, time spend:0.5189ms.
41. [ INFO ] 9. Process reault  success, time spend:0.4425ms.
42. [ INFO ] The result save to E:\GitSpace\OpenVINO-CSharp-API-Samples\model_samples\ppyoloe\ppyoloe_opencvsharp\bin\Release\net6.0\model\test_image_result.jpg

复制代码

     图像预测结果如下图所示：

4. 代码展示

     以下为嘛中所使用的命名空间代码：

1. using OpenCvSharp.Dnn;
2. using OpenCvSharp;
3. using OpenVinoSharp;
4. using OpenVinoSharp.Extensions;
5. using OpenVinoSharp.Extensions.utility;
6. using System.Runtime.InteropServices;
7. using OpenVinoSharp.preprocess;
8. using OpenVinoSharp.Extensions.model;
9. using OpenVinoSharp.Extensions.result;
10. using OpenVinoSharp.Extensions.process;
12. namespace ppyoloe_opencvsharp
13. {
14.     internal class Program
15.     {  
16.             ....
17.     }
18. }

复制代码

     下面为定义的模型预测代码：

• 一般预测流程：
  

1. static void ppyoloe_det(string model_path, string image_path, string device)
2. {
3.     // -------- Step 1. Initialize OpenVINO Runtime Core --------
4.     Core core = new Core();
5.     // -------- Step 2. Read inference model --------
6.     Model model = core.read_model(model_path);
7.     OvExtensions.printf_model_info(model);
8.     // -------- Step 3. Loading a model to the device --------
9.     CompiledModel compiled_model = core.compile_model(model, device);
10.     // -------- Step 4. Create an infer request --------
11.     InferRequest infer_request = compiled_model.create_infer_request();
12.     // -------- Step 5. Process input images --------
13.     Mat image = new Mat(image_path); // Read image by opencvsharp
14.     float[] factor = new float[] { 640.0f / (float)image.Rows, 640.0f / (float)image.Cols };
15.     float[] im_shape = new float[] { 640.0f, 640.0f };
16.     Mat input_mat = CvDnn.BlobFromImage(image, 1.0 / 255.0, new OpenCvSharp.Size(640, 640), 0, true, false);
17.     float[] input_data = new float[640 * 640 * 3];
18.     Marshal.Copy(input_mat.Ptr(0), input_data, 0, input_data.Length);
19.     // -------- Step 6. Set up input data --------
20.     Tensor input_tensor_data = infer_request.get_tensor("image");
21.     input_tensor_data.set_shape(new Shape(1, 3, 640, 640));
22.     input_tensor_data.set_data<float>(input_data);
23.     Tensor input_tensor_factor = infer_request.get_tensor("scale_factor");
24.     input_tensor_factor.set_shape(new Shape(1, 2));
25.     input_tensor_factor.set_data<float>(factor);
26.     // -------- Step 7. Do inference synchronously --------
27.     infer_request.infer();
28.     // -------- Step 8. Get infer result data --------
29.     Tensor output_tensor = infer_request.get_output_tensor(0);
30.     int output_length = (int)output_tensor.get_size();
31.     float[] output_data = output_tensor.get_data<float>(output_length);
32.     // -------- Step 9. Process reault  --------
33.     List<Rect> position_boxes = new List<Rect>();
34.     List<int> class_ids = new List<int>();
35.     List<float> confidences = new List<float>();
36.     for (int i = 0; i < 300; ++i)
37.     {
38.         if (output_data[6 * i + 1] > 0.5)
39.         {
40.             class_ids.Add((int)output_data[6 * i]);
41.             confidences.Add(output_data[6 * i + 1]);
42.             position_boxes.Add(new Rect((int)output_data[6 * i + 2], (int)output_data[6 * i + 3],
43.                 (int)(output_data[6 * i + 4] - output_data[6 * i + 2]),
44.                 (int)(output_data[6 * i + 5] - output_data[6 * i + 3])));
45.         }
46.     }
47.     for (int index = 0; index < class_ids.Count; index++)
48.     {
49.         Cv2.Rectangle(image, position_boxes[index], new Scalar(0, 0, 255), 2, LineTypes.Link8);
50.         Cv2.Rectangle(image, new OpenCvSharp.Point(position_boxes[index].TopLeft.X, position_boxes[index].TopLeft.Y + 30),
51.             new OpenCvSharp.Point(position_boxes[index].BottomRight.X, position_boxes[index].TopLeft.Y), new Scalar(0, 255, 255), -1);
52.         Cv2.PutText(image, class_ids[index] + "-" + confidences[index].ToString("0.00"),
53.             new OpenCvSharp.Point(position_boxes[index].X, position_boxes[index].Y + 25),
54.             HersheyFonts.HersheySimplex, 0.8, new Scalar(0, 0, 0), 2);
55.     }
56.     string output_path = Path.Combine(Path.GetDirectoryName(Path.GetFullPath(image_path)),
57.         Path.GetFileNameWithoutExtension(image_path) + "_result.jpg");
58.     Cv2.ImWrite(output_path, image);
59.     Slog.INFO("The result save to " + output_path);
60.     Cv2.ImShow("Result", image);
61.     Cv2.WaitKey(0);
62. }

复制代码

• 编译预处理步骤到模型方式推理模型：
  

1. static void ppyoloe_det_with_process(string model_path, string image_path, string device)
2. {
3.     // -------- Step 1. Initialize OpenVINO Runtime Core --------
4.     Core core = new Core();
5.     // -------- Step 2. Read inference model --------
6.     Model model = core.read_model(model_path);
7.     OvExtensions.printf_model_info(model);
8.     PrePostProcessor processor = new PrePostProcessor(model);
9.     Tensor input_tensor_pro = new Tensor(new OvType(ElementType.U8), new Shape(1, 640, 640, 3));
10.     InputInfo input_info = processor.input("image");
11.     InputTensorInfo input_tensor_info = input_info.tensor();
12.     input_tensor_info.set_from(input_tensor_pro).set_layout(new Layout("NHWC")).set_color_format(ColorFormat.BGR);
13.     PreProcessSteps process_steps = input_info.preprocess();
14.     process_steps.convert_color(ColorFormat.RGB).resize(ResizeAlgorithm.RESIZE_LINEAR)
15.         .convert_element_type(new OvType(ElementType.F32)).scale(255.0f).convert_layout(new Layout("NCHW"));
16.     Model new_model = processor.build();
17.     // -------- Step 3. Loading a model to the device --------
18.     CompiledModel compiled_model = core.compile_model(new_model, device);
19.     // -------- Step 4. Create an infer request --------
20.     InferRequest infer_request = compiled_model.create_infer_request();
21.     // -------- Step 5. Process input images --------
22.     Mat image = new Mat(image_path); // Read image by opencvsharp
23.     Mat input_image = new Mat();
24.     Cv2.Resize(image, input_image, new OpenCvSharp.Size(640, 640));
25.     float[] factor = new float[] { 640.0f / (float)image.Rows, 640.0f / (float)image.Cols };
26.     float[] im_shape = new float[] { 640.0f, 640.0f };
27.     // -------- Step 6. Set up input data --------
28.     Tensor input_tensor_data = infer_request.get_tensor("image");
29.     byte[] input_data = new byte[3 * 640 * 640];
30.     Marshal.Copy(input_image.Ptr(0), input_data, 0, input_data.Length);
31.     IntPtr destination = input_tensor_data.data();
32.     Marshal.Copy(input_data, 0, destination, input_data.Length);
33.     Tensor input_tensor_factor = infer_request.get_tensor("scale_factor");
34.     input_tensor_factor.set_shape(new Shape(1, 2));
35.     input_tensor_factor.set_data<float>(factor);
36.     // -------- Step 7. Do inference synchronously --------
37.     infer_request.infer();
38.     // -------- Step 8. Get infer result data --------
39.     Tensor output_tensor = infer_request.get_output_tensor(0);
40.     int output_length = (int)output_tensor.get_size();
41.     float[] output_data = output_tensor.get_data<float>(output_length);
42.     // -------- Step 9. Process reault  --------
43.     List<Rect> position_boxes = new List<Rect>();
44.     List<int> class_ids = new List<int>();
45.     List<float> confidences = new List<float>();
47.     for (int i = 0; i < 300; ++i)
48.     {
49.         if (output_data[6 * i + 1] > 0.5)
50.         {
51.             class_ids.Add((int)output_data[6 * i]);
52.             confidences.Add(output_data[6 * i + 1]);
53.             position_boxes.Add(new Rect((int)output_data[6 * i + 2], (int)output_data[6 * i + 3],
54.                 (int)(output_data[6 * i + 4] - output_data[6 * i + 2]),
55.                 (int)(output_data[6 * i + 5] - output_data[6 * i + 3])));
56.         }
57.     }
58.     for (int index = 0; index < class_ids.Count; index++)
59.     {
60.         Cv2.Rectangle(image, position_boxes[index], new Scalar(0, 0, 255), 2, LineTypes.Link8);
61.         Cv2.Rectangle(image, new OpenCvSharp.Point(position_boxes[index].TopLeft.X, position_boxes[index].TopLeft.Y + 30),
62.             new OpenCvSharp.Point(position_boxes[index].BottomRight.X, position_boxes[index].TopLeft.Y), new Scalar(0, 255, 255), -1);
63.         Cv2.PutText(image, class_ids[index] + "-" + confidences[index].ToString("0.00"),
64.             new OpenCvSharp.Point(position_boxes[index].X, position_boxes[index].Y + 25),
65.             HersheyFonts.HersheySimplex, 0.8, new Scalar(0, 0, 0), 2);
66.     }
67.     string output_path = Path.Combine(Path.GetDirectoryName(Path.GetFullPath(image_path)),
68.         Path.GetFileNameWithoutExtension(image_path) + "_result.jpg");
69.     Cv2.ImWrite(output_path, image);
70.     Slog.INFO("The result save to " + output_path);
71.     Cv2.ImShow("Result", image);
72.     Cv2.WaitKey(0);
73. }

复制代码

• 使用封装的方法：
  

1. static void ppyoloe_det_using_extensions(string model_path, string image_path, string device)
2. {
3.     PPYoloeConfig config = new PPYoloeConfig();
4.     config.set_model(model_path);
5.     PPYoloeDet det = new PPYoloeDet(config);
6.     Mat image = Cv2.ImRead(image_path);
7.     DetResult result = det.predict(image);
8.     Mat result_im = Visualize.draw_det_result(result, image);
9.     Cv2.ImShow("Result", result_im);
10.     Cv2.WaitKey(0);
11. }

复制代码

     下面为程序运行的主函数代码，该代码会下载转换好的预测模型，并调用预测方法进行预测：

1. static void Main(string[] args)
2. {
3.     string model_path = "";
4.     string image_path = "";
5.     string device = "CPU";
6.     if (args.Length == 0)
7.     {
8.         if (!Directory.Exists("./model"))
9.         {
10.             Directory.CreateDirectory("./model");
11.         }
12.         if (!File.Exists("./model/model.pdiparams")
13.             && !File.Exists("./model/model.pdmodel"))
14.         {
15.             if (!File.Exists("./model/ppyoloe_plus_crn_l_80e_coco.tar"))
16.             {
17.                 _ = Download.download_file_async("https://github.com/guojin-yan/OpenVINO-CSharp-API-Samples/releases/download/Model/ppyoloe_plus_crn_l_80e_coco.tar",
18.                     "./model/ppyoloe_plus_crn_l_80e_coco.tar").Result;
19.             }
20.             Download.unzip("./model/ppyoloe_plus_crn_l_80e_coco.tar", "./model/");
21.         }
23.         if (!File.Exists("./model/test_image.jpg"))
24.         {
25.             _ = Download.download_file_async("https://github.com/guojin-yan/OpenVINO-CSharp-API-Samples/releases/download/Image/test_det_02.jpg",
26.                 "./model/test_image.jpg").Result;
27.         }
28.         model_path = "./model/model.pdmodel";
29.         image_path = "./model/test_image.jpg";
30.     }
31.     else if (args.Length >= 2)
32.     {
33.         model_path = args[0];
34.         image_path = args[1];
35.         device = args[2];
36.     }
37.     else
38.     {
39.         Console.WriteLine("Please enter the correct command parameters, for example:");
40.         Console.WriteLine("> 1. dotnet run");
41.         Console.WriteLine("> 2. dotnet run <model path> <image path> <device name>");
42.     }
43.     // -------- Get OpenVINO runtime version --------
45.     OpenVinoSharp.Version version = Ov.get_openvino_version();
47.     Slog.INFO("---- OpenVINO INFO----");
48.     Slog.INFO("Description : " + version.description);
49.     Slog.INFO("Build number: " + version.buildNumber);
51.     Slog.INFO("Predict model files: " + model_path);
52.     Slog.INFO("Predict image  files: " + image_path);
53.     Slog.INFO("Inference device: " + device);
54.     Slog.INFO("Start RT-DETR model inference.");
56.     ppyoloe_det(model_path, image_path, device);
57.     //ppyoloe_det_with_process(model_path, image_path, device);
58.     //ppyoloe_det_using_extensions(model_path, image_path, device);
59. }

复制代码

5. 总结

     在该项目中，我们结合之前开发的 OpenVINO™ C# API 项目部署PP-YOLOE模型，实现物体检测。

• 项目完整代码链接为：
  

1. https://github.com/guojin-yan/OpenVINO-CSharp-API-Samples/blob/master/model_samples/ppyoloe/ppyoloe_opencvsharp/Program.cs

复制代码

• 为了方便EmguCV用户使用需求，同时开发了EmguCV版本，项目链接为：
  

1. https://github.com/guojin-yan/OpenVINO-CSharp-API-Samples/blob/master/model_samples/ppyoloe/ppyoloe_emgucv/Program.cs

复制代码

最后如果各位开发者在使用中有任何问题，欢迎大家与我联系。

来源:https://www.cnblogs.com/guojin-blogs/p/18188742
免责声明：由于采集信息均来自互联网，如果侵犯了您的权益，请联系我们【E-Mail:cb@itdo.tech】 我们会及时删除侵权内容，谢谢合作！