OpenCVSharp DNN - 使用Caffe框架模型

用OpenCVSharp （4.5）跑一遍 OpenCV 官方教程。

原 OpenCV 官方教程链接：OpenCV: Load Caffe framework models

using System;
using OpenCvSharp;
using System.Linq;
using System.Text;
using System.Threading.Tasks;
using OpenCvSharp.Dnn;
using System.IO;

namespace ConsoleApp1
{
    class tutorial57 : ITutorial
    {
        public void Run()
        {

            int rszWidth = 224;
            int rszHeight = 244;
            float scale = 1;
            Scalar mean = new Scalar(104, 117, 123);
            Scalar std = new Scalar(0.0, 0.0, 0.0);
            bool swapRB = true;
            bool crop = false;
            int inpWidth = 224;
            int inpHeight = 161;
            string model = @"I:\csharp\images\caffe\bvlc_googlenet.caffemodel";
            string config = @"I:\csharp\images\caffe\bvlc_googlenet.prototxt";
            string classesfile = @"I:\csharp\images\caffe\classification_classes_ILSVRC2012.txt";
            //            string imagefile = @"I:\csharp\images\caffe\space_shuttle.jpg";
            //            string imagefile = @"I:\csharp\images\caffe\bird.jpg";
            //            string imagefile = @"I:\csharp\images\caffe\butterfly.jpg";
            string imagefile = @"I:\csharp\images\caffe\fish.jpg";

            int backendId = 0; //auto
            int targetId = 0; //cpu

            // Open file with classes names.                        
            var classes = File.ReadAllLines(classesfile)
               .Select(line => line.Split(' ').Last())
               .ToArray();

            Net net = Net.ReadNet(model, config);
            net.SetPreferableBackend(Backend.OPENCV);
            net.SetPreferableTarget(Target.CPU);

            // Create a window
            string kWinName = "Deep learning image classification in OpenCV";
            Cv2.NamedWindow(kWinName, WindowFlags.Normal);
            Mat frame = new Mat(imagefile), blob = new Mat();
            if (frame.Empty())
            {
                return;
            }
            if (rszWidth != 0 && rszHeight != 0)
            {
                Cv2.Resize(frame, frame, new Size(rszWidth, rszHeight));
                inpHeight = rszHeight;
                inpWidth = rszWidth;
            }
            blob = CvDnn.BlobFromImage(frame, scale, new Size(inpWidth, inpHeight), mean, swapRB, crop);
            // Check std values.
            if (std.Val0 != 0.0 && std.Val1 != 0.0 && std.Val2 != 0.0)
            {
                // Divide blob by std.
                Cv2.Divide(blob, std, blob);
            }
            net.SetInput(blob);
            Mat prob = net.Forward();
            Point classIdPoint, minLoc;
            double confidence, minValue;
            Cv2.MinMaxLoc(prob.Reshape(1, 1), out minValue, out confidence, out minLoc, out classIdPoint);
            int classId = classIdPoint.X;
            // Put efficiency information.
            double[] layersTimes;
            double freq = Cv2.GetTickFrequency() / 1000;
            double t = net.GetPerfProfile(out layersTimes) / freq;
            String label = String.Format("Inference time:{0:00.0}ms", t);
            Cv2.PutText(frame, label, new Point(0, 15), HersheyFonts.HersheySimplex, 0.5, new Scalar(0, 0, 255));
            // Print predicted class.
            label = String.Format("{0}:{1:0.000}", classes[classId], confidence);
            Cv2.PutText(frame, label, new Point(0, 40), HersheyFonts.HersheySimplex, 0.5, new Scalar(0, 0, 255));
            Cv2.ImShow(kWinName, frame);
            Cv2.WaitKey();
        }
    }
}