.Net使用第三方onnx或ModelBuilder轻松接入AI模型

 <Grid>
        <Button Click="Button_Click" Content="预测一个奥特曼" />
    </Grid>

private void Button_Click(object sender, RoutedEventArgs e) {
            OpenFileDialog dialog= new OpenFileDialog();
            if (dialog.ShowDialog().Value) {
                //Load sample data
                var imageBytes = File.ReadAllBytes(dialog.FileName);
                UltraMan.ModelInput sampleData = new UltraMan.ModelInput() {
                    ImageSource = imageBytes,
                };

                //Load model and predict output
                var result = UltraMan.Predict(sampleData);
                if (result.Score.Any(s=>s>=0.6)) {
                    MessageBox.Show(result.PredictedLabel);
                } else {
                    MessageBox.Show("没有识别到奥特曼");
                }
            }
            
        }

    public class EmotionInput {
            [ImageType(64,64)]
            public MLImage Image { get; set; }  
        }

    public class EmotionOutput {
            [ColumnName("Plus692_Output_0")]
            public float[] Result { get; set; }
        }

     public class EmotionPrediction {
            private readonly string modelFile = "emotion-ferplus-8.onnx";
            private string[] emotions = new string[] { "一般", "快乐", "惊讶", "伤心", "生气", "疑惑", "害怕", "蔑视" };
            private PredictionEngine<EmotionInput, EmotionOutput> predictionEngine;
            public EmotionPrediction()
            {
                MLContext context = new MLContext();
                var emptyData = new List<EmotionInput>();
                var data = context.Data.LoadFromEnumerable(emptyData);
                var pipeline = context.Transforms.ResizeImages("resize", 64, 64, inputColumnName: nameof(EmotionInput.Image), Microsoft.ML.Transforms.Image.ImageResizingEstimator.ResizingKind.Fill).
                    Append(context.Transforms.ExtractPixels("Input3", "resize", Microsoft.ML.Transforms.Image.ImagePixelExtractingEstimator.ColorBits.Blue)).
                    Append(context.Transforms.ApplyOnnxModel(modelFile));
    
                var model = pipeline.Fit(data);
                predictionEngine = context.Model.CreatePredictionEngine<EmotionInput, EmotionOutput>(model);    
            }
    
            public string Predict(string path) {
                using (var stream = new FileStream(path, FileMode.Open)) {
                    using(var bitmap = MLImage.CreateFromStream(stream)) {
                        var result = predictionEngine.Predict(new EmotionInput() { Image = bitmap });
                        var max = result.Result.Max();
                        var index = result.Result.ToList().IndexOf(max);
                        return emotions[index];
                    }
                }
            }
        }

<Grid>
        <Button Click="Button_Click" Content="预测表情" />
    </Grid>

private EmotionPrediction prediction = new EmotionPrediction();

        private void Button_Click(object sender, RoutedEventArgs e) {
            OpenFileDialog dialog = new OpenFileDialog();
            if (dialog.ShowDialog().Value) {
                var result = prediction.Predict(dialog.FileName);
                MessageBox.Show(result);
            }
        }

    public class RTFInput {
            [ImageType(640, 480)]
            public MLImage Image { get; set; }
        }

    public class RTFOutput {
            [ColumnName("scores")]
            [VectorType(1, 17640, 2)]
            public float[] Scores { get; set; }
            [ColumnName("boxes")]
            [VectorType(1, 17640, 4)]
            public float[] Boxes { get; set; }
        }

    public class RTFPrediction {
            private readonly string modelFile = "version-RFB-640.onnx";
            private PredictionEngine<RTFInput, RTFOutput> predictionEngine = null;
            public RTFPrediction() {
                MLContext context = new MLContext();
                var emptyData = new List<RTFInput>();
                var data = context.Data.LoadFromEnumerable(emptyData);
                var pipeline =
                      context.Transforms.ResizeImages(
                      resizing: Microsoft.ML.Transforms.Image.ImageResizingEstimator.ResizingKind.Fill,//填充Resize
                      outputColumnName: "resize",//Resize的结果放置到 data列
                      imageWidth: 640,
                      imageHeight: 480,
                      inputColumnName: nameof(RTFInput.Image)//从Image属性来源,
                      )
                      .Append(
                          context.Transforms.ExtractPixels(
                              offsetImage: 127f,
                              scaleImage: 1 / 128f,
                              inputColumnName: "resize",
                              outputColumnName: "input")
                      ).Append(
                      context.Transforms.ApplyOnnxModel(
                          modelFile: modelFile,
                          inputColumnNames: new string[] { "input" },
                          outputColumnNames: new string[] { "scores", "boxes" }));
                var model = pipeline.Fit(data);
                predictionEngine = context.Model.CreatePredictionEngine<RTFInput, RTFOutput>(model);//生成预测引擎
            }
    
            public ImageSource Predict(string path) {
                using (var stream = new FileStream(path, FileMode.Open)) {
                    using (var bitmap = MLImage.CreateFromStream(stream)) {
                        var prediction = predictionEngine.Predict(new RTFInput() { Image = bitmap });
                        var boxes = ParseBox(prediction);
                        boxes = boxes.Where(b => b.Score > 0.9).OrderByDescending(s => s.Score).ToList();
                        boxes = HardNMS(boxes, 0.4);
                        var bitmapimage = new BitmapImage(new Uri(path));
                        var rtb = new RenderTargetBitmap(bitmap.Width, bitmap.Height, 96, 96, PixelFormats.Pbgra32);
                        int t = 0;
    
                        var dv = new DrawingVisual();
                        using (DrawingContext dc = dv.RenderOpen()) {
                            dc.DrawImage(bitmapimage, new Rect(0, 0, bitmap.Width, bitmap.Height));
                            foreach (var item in boxes) {
                                dc.DrawRectangle(null, new Pen(Brushes.Red, 2), new Rect(item.Rect.X * bitmap.Width, item.Rect.Y * bitmap.Height, item.Rect.Width * bitmap.Width, item.Rect.Height * bitmap.Height));
                            }
                        }
                        rtb.Render(dv);
                        return rtb;
    
                    }
    
                }
            }
    
            private List<Box> ParseBox(RTFOutput prediction) {
                var length = prediction.Boxes.Length / 4;
                var boxes = Enumerable.Range(0, length).Select(i => new Box() {
                    X1 = prediction.Boxes[i * 4],
                    Y1 = prediction.Boxes[i * 4 + 1],
                    X2 = prediction.Boxes[i * 4 + 2],
                    Y2 = prediction.Boxes[i * 4 + 3],
                    Score = prediction.Scores[i * 2 + 1]
                }
                );
                boxes = boxes.OrderByDescending(b => b.Score);
                return boxes.ToList();
            }
    
            public List<Box> HardNMS(List<Box> boxes, double overlapThreshold) {
    
                var selectedBoxes = new List<Box>();
                while (boxes.Count > 0) {
                    // 取出置信度最高的bbox
                    var currentBox = boxes[0];
                    selectedBoxes.Add(currentBox);
    
                    // 计算当前bbox和其余bbox之间的IOU
                    boxes.RemoveAt(0);
                    for (int i = boxes.Count - 1; i >= 0; i--) {
                        var iou = CalculateIOU(currentBox, boxes[i]);
                        if (iou >= overlapThreshold) {
                            boxes.RemoveAt(i);
                        }
                    }
                }
                return selectedBoxes;
            }
            public double CalculateIOU(Box boxA, Box boxB) {
                // 计算相交部分的坐标信息
                float xOverlap = Math.Max(0, Math.Min(boxA.X2, boxB.X2) - Math.Max(boxA.X1, boxB.X1) + 1);
                float yOverlap = Math.Max(0, Math.Min(boxA.Y2, boxB.Y2) - Math.Max(boxA.Y1, boxB.Y1) + 1);
    
                // 计算相交部分的面积和并集部分的面积
                float intersectionArea = xOverlap * yOverlap;
                float unionArea = boxA.Area + boxB.Area - intersectionArea;
    
                // 计算IoU
                double iou = (double)intersectionArea / unionArea;
                return iou;
            }
        }
    

    public class Box {
            public float X1 { get; set; }
            public float Y1 { get; set; }
            public float X2 { get; set; }
            public float Y2 { get; set; }
            public float Score { get; set; }
            // 计算面积
            public float Area => (X2 - X1 + 1) * (Y2 - Y1 + 1);
    
            private Rect GetRect() {
                var hei = Y2 - Y1;
                var wid = X2 - X1;
                if (wid < 0) {
                    wid = 0;
                }
                if (hei < 0) {
                    hei = 0;
                }
                return new Rect(X1, Y1, wid, hei);
            }
    
            private Rect rect = Rect.Empty;
    
            public Rect Rect {
                get {
                    if (rect.IsEmpty) {
                        rect = GetRect();
                    }
                    return rect;
                }
            }
    
        }

<Grid>
        <Grid.RowDefinitions>
            <RowDefinition Height="50" />
            <RowDefinition />
        </Grid.RowDefinitions>
        <Button Click="Button_Click" Content="识别人脸框" />
        <Image x:Name="image" Grid.Row="1" />
    </Grid>

private RTFPrediction prediction = new RTFPrediction();

        private void Button_Click(object sender, RoutedEventArgs e) {
            OpenFileDialog dialog = new OpenFileDialog();
            if (dialog.ShowDialog().Value) {
                var result = prediction.Predict(dialog.FileName);
                if (result!=null) {
                    image.Source = result;
                }
            }
        }