add adjust_gamma

inference_test.py

@@ -93,11 +93,16 @@ class TextProcessor:
             self.vocab[idx] if 1 <= idx <= len(self.vocab) else self.unk_token
             for idx in indices
         )
-        almost_done = self._remove_specials(raw_text)
-        final_text = post_process(almost_done)
+        final_text = self._remove_specials(raw_text)
+        
         return final_text
 
 
+def adjust_gamma(image, gamma=1.0):
+    invGamma = 1.0 / gamma
+    table = np.array([((i / 255.0) ** invGamma) * 255 for i in np.arange(0, 256)]).astype("uint8")
+    return cv2.LUT(image, table)
+    
 
 
 
@@ -106,13 +111,13 @@ def preprocess_image(image_path):
     gray = cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)
 
     # Convert grayscale to 3-channel RGB by duplicating the gray channel
-    img = cv2.cvtColor(gray, cv2.COLOR_GRAY2RGB)
+    img = adjust_gamma(cv2.cvtColor(gray, cv2.COLOR_GRAY2RGB),gamma=2.2)
     mat_in = ncnn.Mat.from_pixels_resize(
         img, ncnn.Mat.PixelType.PIXEL_RGB, img.shape[1], img.shape[0], 224, 224
     )
     
     mean_vals = [0.5, 0.5, 0.5] * 1
-    norm_vals = [1/255, 1/255, 1/255] * 1
+    norm_vals = [1/255, 1/255, 1/255] * 3
     mat_in.substract_mean_normalize(mean_vals, norm_vals)
     return mat_in
 
@@ -145,12 +150,6 @@ def decode_tokens(logits, tokenizer_path="model"):
     return tkz.decode(token_ids, skip_special_tokens=True)
 
 
-def post_process(text):
-    text = "".join(text.split()).replace("…", "...")
-    text = re.sub("[・.]{2,}", lambda x: (x.end() - x.start()) * ".", text)
-    return jaconv.h2z(text, ascii=True, digit=True)
-
-
 def main(image_path):
     mat_in = preprocess_image(image_path)
     logits = run_inference(mat_in)