improve accuracy

we dont need 3 channels for manga, and in my tests the accuracy was greatly improved by lowering the number of channels to only two (maybe B&W)

this can be related to the fact that some files in the model folder do mention that only two channels are supposed to be used.

inference_test.py

@@ -85,7 +85,7 @@ def preprocess_image(image_path):
     mat_in = ncnn.Mat.from_pixels_resize(
         img, ncnn.Mat.PixelType.PIXEL_RGB, img.shape[1], img.shape[0], 224, 224
     )
-    mean_vals, norm_vals = [0.5] * 3, [1 / 255.0] * 3
+    mean_vals, norm_vals = [0.5] * 2, [1 / 255.0] * 2
     mat_in.substract_mean_normalize(mean_vals, norm_vals)
     return mat_in