ONNXRUNTIME C#部署运行YOLO12对象检测

OpenCV学堂

发布于 2026-04-02 17:44:59

1050

文章被收录于专栏：贾志刚-OpenCV学堂贾志刚-OpenCV学堂

ONNXRUNTIME介绍

ONNXRUNTIME是微软开源的深度学习跨平台的推理与加速部署框架，支持CNN、VIT、CLIP、GenAI等模型推理与部署。官方的说明如下：

安装与配置

在VS2022中安装非常简单，直接通过IDE安装即可：

命令行安装方式如下：

dotnet add package Microsoft.ML.OnnxRuntime --version 1.16.0
dotnet add package System.Numerics.Tensors --version 0.1.0

SDK与推理流程

创建推理会话与GPU支持

int gpuDeviceId = 0; // The GPU device ID to execute on
using var gpuSessionOptoins = SessionOptions.MakeSessionOptionWithCudaProvider(gpuDeviceId);
using var session = new InferenceSession("model.onnx", gpuSessionOptoins);

CPU推理会话

using InferenceSession session = new InferenceSession(modelBPath);

创建输入数据

float[] inputData = { 1, 2, 3, 4 };
long[] inputShape = { 1, 4 };
using var inputOrtValue = OrtValue.CreateTensorValueFromMemory(inputData, inputShape);

模型推理

using var runOptions = new RunOptions();
var outputs1 = session.Run(runOptions, inputs1, session1.OutputNames)

推理后处理

// get intermediate value
var outputToFeed = outputs1.First();
// modify the name of the ONNX value
// create input list for session2
var inputs2 = new Dictionary
{
    { "inputNameForModelB", outputToFeed }
};

代码演示

基于YOLO12对象检测的ONNXRUNTIME的C#版本推理演示测试代码如下：

// Setup inputs
var inputs = new List<NamedOnnxValue>
{
    NamedOnnxValue.CreateFromTensor("images", input)
};
var outputs = session_run.Run(inputs);
// from data array to create opencv mat object
var rows = outputs.First().AsEnumerable<float>().ToArray();
List<int> classIds = new List<int>();
List<float> confidences = new List<float>();
List<Rect> boxes = new List<Rect>();
double xFactor = img_w / 640.0;
double yFactor = img_h / 640.0;
for (int r = 0; r < rows.Rows; r++)
{
    Mat row = rows.Row(r);
    //Console.WriteLine("row width: " + row.Rows);
    Mat classesScores = new Mat();
    Rect roi = new Rect(4, 0, 80, 1);
    row[roi].CopyTo(classesScores);
    double minv = 0, maxv = 0;
    Point minLoc = new Point(0, 0);
    Point maxLoc = new Point(0, 0);
    Cv2.MinMaxLoc(classesScores, out minv, out maxv, out minLoc, out maxLoc);
    int classId = maxLoc.X;
    if (classesScores.At<float>(0, classId) > 0.25)
    {
        confidences.Add(classesScores.At<float>(0, classId));
        classIds.Add(classId);
        float cx = row.At<float>(0, 0);
        float cy = row.At<float>(0, 1);
        float w = row.At<float>(0, 2);
        float h = row.At<float>(0, 3);
        int left = (int)((cx - 0.5 * w) * xFactor);
        int top = (int)((cy - 0.5 * h) * yFactor);
        int width = (int)(w * xFactor);
        int height = (int)(h * yFactor);
        boxes.Add(new Rect(left, top, width, height));
    }
}
// Non-maximum suppression
int[] indexes;
CvDnn.NMSBoxes(boxes, confidences, 0.25f, 0.45f, out indexes);
foreach (int index in indexes)
{
    Rect box = boxes[index];
    var color = colors[(classIds[index] % colors.Count)];
    Cv2.Rectangle(frame, box, new Scalar(color.Item1, color.Item2, color.Item3), 2);
    Cv2.Rectangle(frame, new Rect(box.X, box.Y - 20, box.Width, 15), new Scalar(color.Item1, color.Item2, color.Item3), -1);
    Cv2.PutText(frame, classList[classIds[index]], new Point(box.X, box.Y - 10), HersheyFonts.HersheySimplex, 0.5, new Scalar(0, 0, 0));
}
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", new OpenCvSharp.Point(20, 40), HersheyFonts.HersheyPlain, 2, new Scalar(255, 0, 255), 2);
// Show the frame
Cv2.ImShow("OpenCV学堂C#版本-YOLO12对象检测 ", frame);
if (Cv2.WaitKey(1) == 27) break;

CPU推理速度：