C# Onnx CenterNet实现目标检测的示例详解

using Microsoft.ML.OnnxRuntime.Tensors;
using Microsoft.ML.OnnxRuntime;
using OpenCvSharp;
using System;
using System.Collections.Generic;
using System.Windows.Forms;
using System.Linq;
using System.Drawing;
using System.IO;
using OpenCvSharp.Dnn;
using System.Text;
using OpenCvSharp.Flann;
 
namespace Onnx_Demo
{
    public partial class frmMain : Form
    {
        public frmMain()
        {
            InitializeComponent();
        }
 
        string fileFilter = "*.*|*.bmp;*.jpg;*.jpeg;*.tiff;*.tiff;*.png";
        string image_path = "";
 
        DateTime dt1 = DateTime.Now;
        DateTime dt2 = DateTime.Now;
 
        float confThreshold = 0.4f;
        float nmsThreshold = 0.5f;
 
        int inpWidth;
        int inpHeight;
 
        Mat image;
 
        string model_path = "";
 
        SessionOptions options;
        InferenceSession onnx_session;
        Tensor<float> input_tensor;
        Tensor<float> input_tensor_scale;
        List<NamedOnnxValue> input_container;
 
        IDisposableReadOnlyCollection<DisposableNamedOnnxValue> result_infer;
        DisposableNamedOnnxValue[] results_onnxvalue;
 
        List<string> class_names;
        int num_class;
 
        StringBuilder sb = new StringBuilder();
 
        float[] mean = { 0.406f, 0.456f, 0.485f };
        float[] std = { 0.225f, 0.224f, 0.229f };
 
        int num_grid_y;
        int num_grid_x;
 
        float sigmoid(float x)
        {
            return (float)(1.0 / (1.0 + Math.Exp(-x)));
        }
 
        private void button1_Click(object sender, EventArgs e)
        {
            OpenFileDialog ofd = new OpenFileDialog();
            ofd.Filter = fileFilter;
            if (ofd.ShowDialog() != DialogResult.OK) return;
 
            pictureBox1.Image = null;
            pictureBox2.Image = null;
            textBox1.Text = "";
 
            image_path = ofd.FileName;
            pictureBox1.Image = new System.Drawing.Bitmap(image_path);
            image = new Mat(image_path);
        }
 
        private void Form1_Load(object sender, EventArgs e)
        {
            // 创建输入容器
            input_container = new List<NamedOnnxValue>();
 
            // 创建输出会话
            options = new SessionOptions();
            options.LogSeverityLevel = OrtLoggingLevel.ORT_LOGGING_LEVEL_INFO;
            options.AppendExecutionProvider_CPU(0);// 设置为CPU上运行
 
            // 创建推理模型类，读取本地模型文件
            model_path = "model/ctdet_coco_dlav0_384.onnx";
 
            inpHeight = 384;
            inpWidth = 384;
 
            num_grid_y = 96;
            num_grid_x = 96;
 
            onnx_session = new InferenceSession(model_path, options);
 
            // 创建输入容器
            input_container = new List<NamedOnnxValue>();
 
            image_path = "test_img/person.jpg";
            pictureBox1.Image = new Bitmap(image_path);
 
            class_names = new List<string>();
            StreamReader sr = new StreamReader("coco.names");
            string line;
            while ((line = sr.ReadLine()) != null)
            {
                class_names.Add(line);
            }
            num_class = class_names.Count();
 
        }
 
        private unsafe void button2_Click(object sender, EventArgs e)
        {
            if (image_path == "")
            {
                return;
            }
            textBox1.Text = "检测中，请稍等……";
            pictureBox2.Image = null;
            sb.Clear();
            System.Windows.Forms.Application.DoEvents();
 
            image = new Mat(image_path);
            //-----------------前处理--------------------------
            Mat dstimg = new Mat();
            Cv2.CvtColor(image, dstimg, ColorConversionCodes.BGR2RGB);
            Cv2.Resize(dstimg, dstimg, new OpenCvSharp.Size(inpWidth, inpHeight));
            Mat[] mv = new Mat[3];
            Cv2.Split(dstimg, out mv);
            for (int i = 0; i < mv.Length; i++)
            {
                mv[i].ConvertTo(mv[i], MatType.CV_32FC1, 1.0 / (255.0 * std[i]), (0.0 - mean[i]) / std[i]);
            }
            Cv2.Merge(mv, dstimg);
 
            int row = dstimg.Rows;
            int col = dstimg.Cols;
            float[] input_tensor_data = new float[1 * 3 * row * col];
            for (int c = 0; c < 3; c++)
            {
                for (int i = 0; i < row; i++)
                {
                    for (int j = 0; j < col; j++)
                    {
                        float pix = ((float*)(dstimg.Ptr(i).ToPointer()))[j * 3 + c];
                        input_tensor_data[c * row * col + i * col + j] = pix;
                    }
                }
            }
 
            input_tensor = new DenseTensor<float>(input_tensor_data, new[] { 1, 3, inpHeight, inpWidth });
            input_container.Add(NamedOnnxValue.CreateFromTensor("input.1", input_tensor));
 
            //-----------------推理--------------------------
            dt1 = DateTime.Now;
            result_infer = onnx_session.Run(input_container);//运行 Inference 并获取结果
            dt2 = DateTime.Now;
 
            //-----------------后处理--------------------------
            results_onnxvalue = result_infer.ToArray();
 
            float ratioh = (float)image.Rows / inpHeight;
            float ratiow = (float)image.Cols / inpWidth;
            float stride = inpHeight / num_grid_y;
 
            float[] pscore = results_onnxvalue[0].AsTensor<float>().ToArray();
            float[] pxy = results_onnxvalue[1].AsTensor<float>().ToArray();
            float[] pwh = results_onnxvalue[2].AsTensor<float>().ToArray();
            int area = num_grid_y * num_grid_x;
 
            List<float> confidences = new List<float>();
            List<Rect> position_boxes = new List<Rect>();
            List<int> class_ids = new List<int>();
            Result result = new Result();
 
            for (int i = 0; i < num_grid_y; i++)
            {
                for (int j = 0; j < num_grid_x; j++)
                {
                    float max_class_score = -1000;
                    int class_id = -1;
                    for (int c = 0; c < num_class; c++)
                    {
                        float score = sigmoid(pscore[c * area + i * num_grid_x + j]);
                        if (score > max_class_score)
                        {
                            max_class_score = score;
                            class_id = c;
                        }
                    }
 
                    if (max_class_score > confThreshold)
                    {
                        float cx = (pxy[i * num_grid_x + j] + j) * stride * ratiow;  ///cx
                        float cy = (pxy[area + i * num_grid_x + j] + i) * stride * ratioh;   ///cy
                        float w = pwh[i * num_grid_x + j] * stride * ratiow;   ///w
                        float h = pwh[area + i * num_grid_x + j] * stride * ratioh;  ///h
 
                        int x = (int)Math.Max(cx - 0.5 * w, 0);
                        int y = (int)Math.Max(cy - 0.5 * h, 0);
                        int width = (int)Math.Min(w, image.Cols - 1);
                        int height = (int)Math.Min(h, image.Rows - 1);
 
                        position_boxes.Add(new Rect(x, y, width, height));
                        class_ids.Add(class_id);
                        confidences.Add(max_class_score);
                    }
                }
            }
 
            // NMS非极大值抑制
            int[] indexes = new int[position_boxes.Count];
            CvDnn.NMSBoxes(position_boxes, confidences, confThreshold, nmsThreshold, out indexes);
 
            for (int i = 0; i < indexes.Length; i++)
            {
                int index = indexes[i];
                result.add(confidences[index], position_boxes[index], class_names[class_ids[index]]);
            }
 
            if (pictureBox2.Image != null)
            {
                pictureBox2.Image.Dispose();
            }
 
            sb.AppendLine("推理耗时:" + (dt2 - dt1).TotalMilliseconds + "ms");
            sb.AppendLine("------------------------------");
 
            // 将识别结果绘制到图片上
            Mat result_image = image.Clone();
            for (int i = 0; i < result.length; i++)
            {
                Cv2.Rectangle(result_image, result.rects[i], new Scalar(0, 0, 255), 2, LineTypes.Link8);
 
                Cv2.Rectangle(result_image, new OpenCvSharp.Point(result.rects[i].TopLeft.X - 1, result.rects[i].TopLeft.Y - 20),
                    new OpenCvSharp.Point(result.rects[i].BottomRight.X, result.rects[i].TopLeft.Y), new Scalar(0, 0, 255), -1);
 
                Cv2.PutText(result_image, result.classes[i] + "-" + result.scores[i].ToString("0.00"),
                    new OpenCvSharp.Point(result.rects[i].X, result.rects[i].Y - 4),
                    HersheyFonts.HersheySimplex, 0.6, new Scalar(0, 0, 0), 1);
 
                sb.AppendLine(string.Format("{0}：{1},({2},{3},{4},{5})"
                    , result.classes[i]
                    , result.scores[i].ToString("0.00")
                    , result.rects[i].TopLeft.X
                    , result.rects[i].TopLeft.Y
                    , result.rects[i].BottomRight.X
                    , result.rects[i].BottomRight.Y
                    ));
            }
 
            textBox1.Text = sb.ToString();
            pictureBox2.Image = new System.Drawing.Bitmap(result_image.ToMemoryStream());
 
            result_image.Dispose();
            dstimg.Dispose();
            image.Dispose();
 
        }
 
        private void pictureBox2_DoubleClick(object sender, EventArgs e)
        {
            Common.ShowNormalImg(pictureBox2.Image);
        }
 
        private void pictureBox1_DoubleClick(object sender, EventArgs e)
        {
            Common.ShowNormalImg(pictureBox1.Image);
        }
    }
}