v2

BetterGenshinImpact.Test/MainWindow.xaml

@@ -78,5 +78,13 @@
                 VerticalAlignment="Center"
                 Click="ZoomOut"
                 Content="缩小地图到256" />
+        <Button Canvas.Left="120"
+                Canvas.Top="40"
+                Width="100"
+                Height="23"
+                HorizontalAlignment="Left"
+                VerticalAlignment="Center"
+                Click="CameraTest"
+                Content="新视角识别测试" />
     </Canvas>
 </Window>

BetterGenshinImpact.Test/MainWindow.xaml.cs

@@ -5,6 +5,9 @@ using BetterGenshinImpact.Test.Simple.AllMap;
 using BetterGenshinImpact.Test.Simple.Track;
 using BetterGenshinImpact.Test.View;
 using System.Windows;
+using BetterGenshinImpact.GameTask.Common.Map;
+using OpenCvSharp;
+using Window = System.Windows.Window;
 
 namespace BetterGenshinImpact.Test;
 
@@ -85,4 +88,10 @@ public partial class MainWindow : Window
     {
         AvatarClassifyTransparentGen.GenAll();
     }
+
+    private void CameraTest(object sender, RoutedEventArgs e)
+    {
+        CameraOrientationV2 cameraOrientation = new();
+        cameraOrientation.PredictRotation(new Mat(@"E:\HuiTask\更好的原神\地图匹配\比较\小地图\Clip_20240323_183119.png"));
+    }
 }

BetterGenshinImpact/BetterGenshinImpact.csproj

@@ -48,6 +48,7 @@
 		<PackageReference Include="Microsoft.ML.OnnxRuntime.DirectML" Version="1.18.1" />
 		<PackageReference Include="Microsoft.ML.OnnxRuntime.Managed" Version="1.18.1" />
 		<PackageReference Include="Microsoft.Web.WebView2" Version="1.0.2592.51" />
+		<PackageReference Include="NumpyDotNet" Version="0.9.86.2" />
 		<PackageReference Include="Ookii.Dialogs.Wpf" Version="5.0.1" />
 		<PackageReference Include="OpenCvSharp4.WpfExtensions" Version="4.8.0.20230708" />
 		<PackageReference Include="OpenCvSharp4.Extensions" Version="4.8.0.20230708" />

BetterGenshinImpact/GameTask/Common/Map/CameraOrientationV2.cs

@@ -0,0 +1,155 @@
+using System;
+using NumpyDotNet;
+using OpenCvSharp;
+
+namespace BetterGenshinImpact.GameTask.Common.Map;
+
+public class CameraOrientationV2
+{
+    private double tplSize = 216;
+    private double tplOutRad = 78;
+    private int tplInnRad = 19;
+    private ndarray alphaParams1 = np.array(new[] { 18.632, 20.157, 24.093, 34.617, 38.566, 41.94, 47.654, 51.087, 58.561, 63.925, 67.759, 71.77, 75.214 });
+
+    private double rLength = 60;
+    private int thetaLength = 360;
+    private int peakWidth;
+    private ndarray rotationRemapDataX;
+    private ndarray rotationRemapDataY;
+
+    private ndarray alphaMask1;
+    private ndarray alphaMask2;
+    private double rotation;
+    private double rotationConfidence;
+
+    public CameraOrientationV2()
+    {
+        this.peakWidth = thetaLength / 4;
+
+        var points = GeneratePoints(tplSize / 2.0, tplSize / 2.0, tplInnRad, tplOutRad, rLength, thetaLength);
+        this.rotationRemapDataX = points.Item1;
+        this.rotationRemapDataY = points.Item2;
+
+        var masks = CreatAlphaMask();
+        this.alphaMask1 = masks.Item1;
+        this.alphaMask2 = masks.Item2;
+    }
+
+    private ndarray ApplyMask(ndarray img, ndarray mask, double bkg)
+    {
+        return (img.astype(np.Float32) - bkg) / mask * 255 + bkg;
+    }
+
+    private ndarray Bgr2h(ndarray bgr)
+    {
+        var cmax = np.max(bgr, axis: -1);
+        var cmin = np.min(bgr, axis: -1);
+        var mask = cmax > cmin;
+
+        var hue = np.zeros_like(cmax).astype(np.Float32);
+        // var maskIndices = np.where(mask);
+
+        hue[mask] = ((cmax[mask] - 2 * (ndarray)cmin[mask] - ((ndarray)bgr["...", 2])[mask] + ((ndarray)bgr["...", 1])[mask] + ((ndarray)bgr["...", 0])[mask]) / ((ndarray)cmax[mask] - (ndarray)cmin[mask])) * 60;
+
+        hue[np.where(~mask)] = -1;
+        hue = (ndarray)np.where((ndarray)bgr["...", 1] >= (ndarray)bgr["...", 0], hue, 360 - hue);
+
+        return hue;
+    }
+
+    private (ndarray, ndarray) GeneratePoints(double x, double y, double d1, double d2, double n, double m)
+    {
+        var r = np.linspace(d1, d2, ref n);
+        var theta = np.linspace(0, 360, ref m, endpoint: false);
+        var thetaRad = np.radians(theta);
+
+        var meshgrid = np.meshgrid([r, thetaRad]);
+        var rGrid = meshgrid[0];
+        var thetaGrid = meshgrid[1];
+
+        var xCartesian = rGrid * np.cos(thetaGrid) + x;
+        var yCartesian = rGrid * np.sin(thetaGrid) + y;
+
+        return (xCartesian.astype(np.Float32), yCartesian.astype(np.Float32));
+    }
+
+    private (ndarray, ndarray) CreatAlphaMask()
+    {
+        var values = np.linspace(tplInnRad, tplOutRad, ref rLength);
+        var alphaMask1 = (229 + np.searchsorted(alphaParams1, values)).astype(np.Float32);
+        var alphaMask2 = (111.7 + 1.836 * values).astype(np.Float32);
+
+        return (alphaMask1, alphaMask2);
+    }
+
+    /// <summary>
+    /// 
+    /// </summary>
+    /// <param name="image">3通道图像</param>
+    /// <param name="confidence"></param>
+    /// <returns></returns>
+    public double? PredictRotation(Mat image, double confidence = 0.3)
+    {
+        Mat remap = new();
+        Cv2.Remap(image, remap, InputArray.Create(rotationRemapDataX.AsFloatArray()), InputArray.Create(rotationRemapDataY.AsFloatArray()),
+            InterpolationFlags.Lanczos4);
+        
+        remap.GetArray(out float[] remapArray);
+        
+        var bgrImg = ApplyMask(np.array(remapArray), alphaMask1.reshape(-1, 1), 0);
+        var hImg = Bgr2h(bgrImg);
+        var faImg = np.mean(bgrImg, axis: 2);
+        var fbImg = ApplyMask(faImg, alphaMask2, 255);
+
+        var histA = new Mat();
+        var histB = new Mat();
+
+        Rangef[] ranges = [new(0, 360), new(0, 256)];
+        int[] histSize = [360, 256];
+        Cv2.CalcHist([InputArray.Create(hImg.AsFloatArray()).GetMat(), InputArray.Create(faImg.AsFloatArray()).GetMat()], [0, 1], null, histA, 2, histSize, ranges);
+        Cv2.CalcHist([InputArray.Create(hImg.AsFloatArray()).GetMat(), InputArray.Create(fbImg.AsFloatArray()).GetMat()], [0, 1], null, histB, 2, histSize, ranges);
+
+        var result = np.zeros((thetaLength, rLength), np.UInt8);
+        var h = np.floor(hImg).astype(np.Int32);
+        var fa = np.floor(faImg).astype(np.Int32);
+        var fb = np.floor(fbImg).astype(np.Int32);
+
+        var flag = (h >= 0) & (h < 360) & (fa >= 0) & (fa < 256) & (fb >= 0) & (fb < 256);
+
+        var histANp = np.array(histA);
+        var histBNp = np.array(histB);
+        result[flag] = np.select(
+            [
+                (ndarray)histANp[h[flag], fa[flag]] > (ndarray)histBNp[h[flag], fb[flag]],
+                ((ndarray)histANp[h[flag], fa[flag]]).Equals((ndarray)histBNp[h[flag], fb[flag]]),
+                (ndarray)histANp[h[flag], fa[flag]] < (ndarray)histBNp[h[flag], fb[flag]]
+            ],
+            [
+                np.array(0),
+                np.array(128),
+                np.array(255)
+            ]
+        ).astype(np.UInt8);
+
+        // 处理结果计算
+        var resultMean = np.mean(result, axis: 1);
+        var resultConv = np.cumsum(resultMean - np.roll(resultMean, peakWidth)) + (double)np.sum((ndarray)resultMean[-peakWidth + ":"]);
+
+        var maxInd = np.argmax(resultConv);
+        var degree = ((int)maxInd + 0.5) / thetaLength * 360 - 45;
+
+        if (degree < 0)
+            degree = 360 + degree;
+
+        this.rotation = degree;
+        this.rotationConfidence = (float)resultConv[(int)maxInd] / (peakWidth * 255);
+
+        if (this.rotationConfidence < confidence)
+        {
+            Console.WriteLine($"置信度{this.rotationConfidence}<{confidence}, 不可靠视角 {this.rotation}");
+            return null;
+        }
+
+        return degree;
+    }
+}