计算机视觉 YOLO V5 ONNX模型在C# 中部署

关于Yolo5训练自定义数据集，请参考此教程：YOLO5训练自定义数据集

完整项目（VS2022） 链接：https://pan.baidu.com/s/17jQUx-Dp0YYC4YWzYLs0AA?pwd=yzj9 提取码：yzj9

一 YOLO 导出ONNX模型

在export.py 中修改参数（如下图），运行导出ONNX

二 Yolov5Net.Scorer.dll文件设置及编译

C# ONNX模型推理dll库： https://github.com/mentalstack/yolov5-net 根据模型修改model文件后，再编译dll，在项目中引用此dll

public class YoloCocoP5Model : YoloModel

{

//输入图像大小

public override int Width { get; set; } = 640;

public override int Height { get; set; } = 640;

public override int Depth { get; set; } = 3;

///

public override int Dimensions { get; set; } = 6;

public override int[] Strides { get; set; } = new int[] { 8, 16, 32 };

public override int[][][] Anchors { get; set; } = new int[][][]

{

new int[][] { new int[] { 010, 13 }, new int[] { 016, 030 }, new int[] { 033, 023 } },

new int[][] { new int[] { 030, 61 }, new int[] { 062, 045 }, new int[] { 059, 119 } },

new int[][] { new int[] { 116, 90 }, new int[] { 156, 198 }, new int[] { 373, 326 } }

};

public override int[] Shapes { get; set; } = new int[] { 80, 40, 20 };

//阈值设定，大于此阈值才识别

public override float Confidence { get; set; } = 0.40f;

public override float MulConfidence { get; set; } = 0.25f;

public override float Overlap { get; set; } = 0.45f;

//ONNX模型的输出变量名称，可在https://netron.app/中查看

public override string[] Outputs { get; set; } = new[] { "output0" };

public override List Labels { get; set; } = new List()

{

//模型中的对象的种类信息

new YoloLabel { Id = 0, Name = "Stop-Sign" },

};

public override bool UseDetect { get; set; } = true;

public YoloCocoP5Model()

{

}

}

三 C# 中模型进行推理

C# 项目编译环境为>=.NET5,本案例使用.NET6需要安装以下库：Microsoft.ML.OnnxRuntime （ONNX运行时）OpenCvSharp4.Extensions （输出图像框定对象使用）OpenCvSharp4.Windows （输出图像框定对象使用）同时需要引用上一步编译的：Yolov5Net.Scorer.dll

测试代码：

using Yolov5Net.Scorer.Models;

using Yolov5Net.Scorer;

using OpenCvSharp;

using System.Windows.Forms;

using Point = OpenCvSharp.Point;

namespace YOLO_WinformDemo

{

///

public partial class Form1 : Form

{

YoloScorer scorer;

string filePath = "";

public Form1()

{

InitializeComponent();

//加载模型地址

scorer = new YoloScorer(Application.StartupPath+ "\\onnx\\best.onnx");

}

private void btn_selectFile_Click(object sender, EventArgs e)

{

OpenFileDialog openFile=new OpenFileDialog();

if(openFile.ShowDialog() == DialogResult.OK)

{

filePath = openFile.FileName;

using var image = Image.FromFile(openFile.FileName);

picbox_Display.BackgroundImage = AddInfoToImage(image);

}

}

///

///

///

public Image AddInfoToImage(Image inputImage)

{

List predictions = scorer.Predict(inputImage);

if (predictions.Count == 0) {

return inputImage;

}

else

{

Mat inputMat = Cv2.ImRead(filePath, ImreadModes.Color);

foreach (YoloPrediction prediction in predictions)

{

Point p1 = new Point(prediction.Rectangle.X, prediction.Rectangle.Y);//矩形左上顶点

Point p2 = new Point(prediction.Rectangle.X+ prediction.Rectangle.Width, prediction.Rectangle.Y + prediction.Rectangle.Height);//矩形右下顶点

Point p3 = new Point(prediction.Rectangle.X, prediction.Rectangle.Y - 60);

Scalar scalar = new Scalar(0, 0, 255);

Cv2.Rectangle(inputMat, p1, p2, scalar, 7);

Cv2.PutText(inputMat, prediction.Label.Name+" "+ Math.Round(prediction.Score,2), p3, HersheyFonts.HersheyDuplex, 2, scalar, 4);

}

return OpenCvSharp.Extensions.BitmapConverter.ToBitmap(inputMat);

}

}

}

}

测试效果：

好文推荐

 您阅读本篇文章共花了：