C# OpenVINO Cls 图像分类

目录

介绍

效果

耗时

项目

代码

下载

---

介绍

OpenVINO.NET github地址：https://github.com/sdcb/OpenVINO.NET

High quality .NET wrapper for OpenVINO™ toolkit.

在AI的应用越来越广泛的今天，优化深度学习模型并进行推理部署已经成为了一门必要的技术。Intel开发的OpenVINO工具包（Open Visual Inference and Neural network Optimization）就是这样一款强大的工具。作为一个开源的工具包，OpenVINO为开发者提供了强大的深度学习模型优化和推理功能，支持跨不同的Intel硬件平台进行部署，包括CPU, 集成GPU, Intel Movidius VPU, 和FPGAs。该工具包的初衷就是实现一处编码后，能在任何地方部署的机器学习推理的解决方案。

效果

耗时

class id=brown_bear, score=0.86
preprocess time: 0.00ms
infer time: 2.72ms
postprocess time: 0.02ms
Total time: 2.74ms

项目

VS2022

.net framework 4.8

OpenCvSharp 4.8

Sdcb.OpenVINO

代码

Model rawModel = OVCore.Shared.ReadModel(model_path);
 PrePostProcessor pp = rawModel.CreatePrePostProcessor();
 PreProcessInputInfo inputInfo = pp.Inputs.Primary;inputInfo.TensorInfo.Layout = Sdcb.OpenVINO.Layout.NHWC;
 inputInfo.ModelInfo.Layout = Sdcb.OpenVINO.Layout.NCHW;Model m = pp.BuildModel();
 CompiledModel cm = OVCore.Shared.CompileModel(m, "CPU");
 InferRequest ir = cm.CreateInferRequest();Shape inputShape = m.Inputs.Primary.Shape;
Stopwatch stopwatch = new Stopwatch();
 Mat resized = src.Resize(new OpenCvSharp.Size(inputShape[2], inputShape[1]));
 Mat f32 = new Mat();
 resized.ConvertTo(f32, MatType.CV_32FC3, 1.0 / 255);using (Tensor input = Tensor.FromRaw(
      new ReadOnlySpan<byte>((void*)f32.Data, (int)((long)f32.DataEnd - (long)f32.DataStart)),
     new Shape(1, f32.Rows, f32.Cols, 3),
     ov_element_type_e.F32))
 {
     ir.Inputs.Primary = input;
 }
 double preprocessTime = stopwatch.Elapsed.TotalMilliseconds;
 stopwatch.Restart();ir.Run();
 double inferTime = stopwatch.Elapsed.TotalMilliseconds;
 stopwatch.Restart();

using OpenCvSharp;
using Sdcb.OpenVINO;
using Sdcb.OpenVINO.Natives;
using System;
using System.Diagnostics;
using System.Drawing;
using System.Text;
using System.Windows.Forms;
using System.Xml.Linq;
using System.Xml.XPath;

namespace OpenVINO_Cls_图像分类
{
    public partial class Form1 : Form
    {
        public Form1()
        {
            InitializeComponent();
        }

        string fileFilter = "*.*|*.bmp;*.jpg;*.jpeg;*.tiff;*.tiff;*.png";
        string image_path = "";
        string startupPath;
        string model_path;
        Mat src;
        string[] dicts;

        StringBuilder sb = new StringBuilder();

        private void button1_Click(object sender, EventArgs e)
        {
            OpenFileDialog ofd = new OpenFileDialog();
            ofd.Filter = fileFilter;
            if (ofd.ShowDialog() != DialogResult.OK) return;
            pictureBox1.Image = null;
            image_path = ofd.FileName;
            pictureBox1.Image = new Bitmap(image_path);
            textBox1.Text = "";
            src = new Mat(image_path);
            pictureBox2.Image = null;
        }

        unsafe private void button2_Click(object sender, EventArgs e)
        {
            if (pictureBox1.Image == null)
            {
                return;
            }

            pictureBox2.Image = null;
            textBox1.Text = "";
            sb.Clear();

            Model rawModel = OVCore.Shared.ReadModel(model_path);
            PrePostProcessor pp = rawModel.CreatePrePostProcessor();
            PreProcessInputInfo inputInfo = pp.Inputs.Primary;

            inputInfo.TensorInfo.Layout = Sdcb.OpenVINO.Layout.NHWC;
            inputInfo.ModelInfo.Layout = Sdcb.OpenVINO.Layout.NCHW;

            Model m = pp.BuildModel();
            CompiledModel cm = OVCore.Shared.CompileModel(m, "CPU");
            InferRequest ir = cm.CreateInferRequest();

            Shape inputShape = m.Inputs.Primary.Shape;

            Stopwatch stopwatch = new Stopwatch();
            Mat resized = src.Resize(new OpenCvSharp.Size(inputShape[2], inputShape[1]));
            Mat f32 = new Mat();
            resized.ConvertTo(f32, MatType.CV_32FC3, 1.0 / 255);

            using (Tensor input = Tensor.FromRaw(
                 new ReadOnlySpan<byte>((void*)f32.Data, (int)((long)f32.DataEnd - (long)f32.DataStart)),
                new Shape(1, f32.Rows, f32.Cols, 3),
                ov_element_type_e.F32))
            {
                ir.Inputs.Primary = input;
            }
            double preprocessTime = stopwatch.Elapsed.TotalMilliseconds;
            stopwatch.Restart();

            ir.Run();
            double inferTime = stopwatch.Elapsed.TotalMilliseconds;
            stopwatch.Restart();

            using (Tensor output = ir.Outputs.Primary)
            {
                ReadOnlySpan<float> data = output.GetData<float>();
                int maxIndex = Common.MaxIndexOfSpan(data);
                double postProcessTime = stopwatch.Elapsed.TotalMilliseconds;
                stopwatch.Stop();

                sb.AppendLine($"class id={dicts[maxIndex]}, score={data[maxIndex]:F2}");
                double totalTime = preprocessTime + inferTime + postProcessTime;
                sb.AppendLine($"preprocess time: {preprocessTime:F2}ms");
                sb.AppendLine($"infer time: {inferTime:F2}ms");
                sb.AppendLine($"postprocess time: {postProcessTime:F2}ms");
                sb.AppendLine($"Total time: {totalTime:F2}ms");

                Mat result_image = src.Clone();
                Cv2.PutText(result_image
                , $"class id={dicts[maxIndex]}, score={data[maxIndex]:F2}"
                , new OpenCvSharp.Point(10, 30)
                , HersheyFonts.HersheySimplex
                , 1
                , new Scalar(0, 0, 255)
                , 2);

                pictureBox2.Image = new Bitmap(result_image.ToMemoryStream());
                textBox1.Text = sb.ToString();
            }
        }


        private void Form1_Load(object sender, EventArgs e)
        {
            startupPath = Application.StartupPath;
            model_path = startupPath + "\\yolov8n-cls.xml";
            dicts = XDocument.Load(model_path)
            .XPathSelectElement(@"/net/rt_info/model_info/labels").Attribute("value").Value
            .Split(' ');
        }
    }
}

下载

可执行程序exe下载

源码下载