This is used a lot of model to solve gastic cancer problem

Image_Process/Image_Generator.py

@@ -16,7 +16,6 @@ class Image_generator():
 
     def Processing_Main(self, DataLoader):
         data_size = 0
-        self.stop = 12
         File = Process_File()
         torch.manual_seed(Training_Config["Random_Seed"])
 
@@ -28,7 +27,7 @@ class Image_generator():
                 產生出資料強化後的影像
             '''
             # data_size = self.Generator_main(DataLoader, 5, data_size) # 執行
-
+            self.stop = len(DataLoader) * 3
             for i in range(1, 6, 1):
                 print(f"\nAugmentation {i} Generator image")
                 data_size = self.Generator_main(DataLoader, i, data_size) # 執行
@@ -119,7 +118,7 @@ class Image_generator():
         if judge == 1:
             return v2.Compose([
                 v2.RandomRotation(30),
-                # v2.RandomResizedCrop(224, scale=(0.8, 1.0)),
+                v2.RandomResizedCrop(224, scale=(0.8, 1.0)),
                 v2.RandomHorizontalFlip(),
                 v2.GaussianBlur(7,3),
                 v2.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1),
@@ -127,14 +126,14 @@ class Image_generator():
         elif judge == 2:
             return v2.Compose([
                 v2.RandomRotation(180),
-                # v2.RandomResizedCrop(224, scale=(0.7, 1.0)),
+                v2.RandomResizedCrop(224, scale=(0.7, 1.0)),
                 v2.RandomHorizontalFlip(),
                 v2.RandomVerticalFlip(),
             ])
         elif judge == 3:
             return v2.Compose([
                 v2.RandomRotation(45, expand = True),
-                # v2.RandomResizedCrop(224, scale=(0.9, 1.0)),
+                v2.RandomResizedCrop(224, scale=(0.9, 1.0)),
                 v2.RandomAffine(degrees=20, shear=0.2),
                 v2.ColorJitter(brightness=0.2, contrast=0.2),
                 v2.RandomHorizontalFlip(),
@@ -143,7 +142,7 @@ class Image_generator():
         elif judge == 4:
             return v2.Compose([
                 v2.RandomRotation(50),
-                # v2.RandomResizedCrop(224, scale=(0.75, 1.0)),
+                v2.RandomResizedCrop(224, scale=(0.75, 1.0)),
                 v2.RandomAffine(degrees=30, shear=0.25),
                 v2.GaussianBlur(7,3),
                 v2.ColorJitter(brightness=0.3, contrast=0.3, saturation=0.3, hue=0.2),

Image_Process/preprocess_training_data.py

@@ -0,0 +1,171 @@
+import cv2
+import numpy as np
+import os
+import sys
+from glob import glob
+from tqdm import tqdm
+
+# Add parent directory to path to import utils
+sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+try:
+    from utils.Stomach_Config import Loading_Config, Training_Config
+except ImportError:
+    print("Warning: Could not import Loading_Config. Using default labels.")
+    Loading_Config = {
+        "Training_Labels": ["stomach_cancer_Crop", "Normal_Crop", "Have_Question_Crop"]
+    }
+    Training_Config = {
+        "Image_Size": 224
+    }
+
+def preprocess_image(image_path, save_path, 
+                     target_size=(224, 224),
+                     blur_ksize=(5, 5), 
+                     adaptive_block_size= 13, 
+                     adaptive_C=4,
+                     gradient_ksize=3,
+                     closing_ksize=5,
+                     min_area=100):
+    """
+    Process image using the modified pipeline:
+    1. Resize (縮放)
+    2. Grayscale (灰階)
+    3. GaussianBlur (3-7)
+    4. Adaptive Threshold (blockSize 11-51 odd, C 2-10, THRESH_BINARY or INV)
+    5. Morphological Gradient (kernel 3x3 - 7x7)
+    6. Closing (Fill holes / Connect edges)
+    7. Connected Components Filter (Remove small components)
+    8. FindContours
+    9. Generate Mask and Apply to Original
+    """
+    
+    # 1. Read Image
+    img = cv2.imread(image_path)
+    if img is None:
+        print(f"Error reading {image_path}")
+        return
+
+    # 2. Resize
+    # Use target_size from config if available, default to 224x224
+    img_resized = cv2.resize(img, target_size)
+
+    # 3. Grayscale
+    if len(img_resized.shape) == 3:
+        gray = cv2.cvtColor(img_resized, cv2.COLOR_BGR2GRAY)
+    else:
+        gray = img_resized
+
+    # 4. GaussianBlur (3-7)
+    blurred = cv2.GaussianBlur(gray, blur_ksize, 0)
+
+    # 5. Adaptive Threshold (blockSize 11-51 odd, C 2-10)
+    thresh = cv2.adaptiveThreshold(blurred, 255, 
+                                   cv2.ADAPTIVE_THRESH_GAUSSIAN_C, 
+                                   cv2.THRESH_BINARY_INV, 
+                                   adaptive_block_size, 
+                                   adaptive_C)
+
+    # 6. Morphological Gradient (kernel 3x3 - 7x7)
+    kernel_grad = cv2.getStructuringElement(cv2.MORPH_RECT, (gradient_ksize, gradient_ksize))
+    gradient = cv2.morphologyEx(thresh, cv2.MORPH_GRADIENT, kernel_grad)
+    
+    # 7. Closing (Fill holes / Connect edges)
+    kernel_close = cv2.getStructuringElement(cv2.MORPH_RECT, (closing_ksize, closing_ksize)) 
+    closed = cv2.morphologyEx(gradient, cv2.MORPH_CLOSE, kernel_close)
+
+    # 8. Connected Components Filter (Remove small components)
+    num_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(closed, connectivity=8)
+    
+    # Create a mask with only large components
+    filtered_components_mask = np.zeros_like(closed)
+    for i in range(1, num_labels): # Skip background (0)
+        area = stats[i, cv2.CC_STAT_AREA]
+        if area >= min_area:
+            filtered_components_mask[labels == i] = 255
+
+    # 9. FindContours (on the filtered components)
+    contours, _ = cv2.findContours(filtered_components_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+    
+    # Create a filled mask from contours
+    final_mask = np.zeros_like(closed)
+    
+    for cnt in contours:
+        # We can also apply aspect ratio/solidity filters here if needed, 
+        # but user specifically asked for "Connected Components Filter -> FindContours"
+        # Assuming CC filter is the main filter now.
+        cv2.drawContours(final_mask, [cnt], -1, 255, -1) # Fill contour
+
+    # 10. Apply Mask to Original Image (Combine)
+    masked_img = cv2.bitwise_and(img_resized, img_resized, mask=final_mask)
+
+    # 11. Save result
+    cv2.imwrite(save_path, masked_img)
+
+def main():
+    base_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) # Pytorch/
+    
+    # Construct paths
+    dataset_path = os.path.abspath(os.path.join(base_dir, "..", "Dataset", "Training"))
+    
+    if not os.path.exists(dataset_path):
+        training_path = os.path.abspath(os.path.join(base_dir, "..", "Dataset", "Training_Crop"))
+        if os.path.exists(training_path):
+            dataset_path = training_path
+            print(f"Dataset/Training_Crop not found. Using {dataset_path}")
+        else:
+             print(f"Error: Neither Training_Crop nor Training directories found at {dataset_path} or {training_path}")
+             return
+
+    # Output directory for the new strategy (Strategy 3)
+    save_root = os.path.abspath(os.path.join(base_dir, "..", "Dataset", "Masked_Training_Crop_Strategy3"))
+    
+    print(f"Base Directory: {base_dir}")
+    print(f"Dataset Path: {dataset_path}")
+    print(f"Save Path: {save_root}")
+    
+    if not os.path.exists(save_root):
+        os.makedirs(save_root)
+        print(f"Created save directory: {save_root}")
+
+    labels = Loading_Config.get("Training_Labels", [])
+    if not labels:
+         labels = ["stomach_cancer_Crop", "Normal_Crop", "Have_Question_Crop"]
+    
+    image_size = Training_Config.get("Image_Size", 224)
+
+    for label in labels:
+        label_dir = os.path.join(dataset_path, label)
+        save_label_dir = os.path.join(save_root, label)
+        
+        if not os.path.exists(save_label_dir):
+            os.makedirs(save_label_dir)
+
+        if not os.path.exists(label_dir):
+            print(f"Skipping {label}: Directory not found at {label_dir}")
+            continue
+
+        # Find images
+        types = ('*.jpg', '*.png', '*.jpeg', '*.bmp')
+        images = []
+        for t in types:
+            images.extend(glob(os.path.join(label_dir, t)))
+        
+        print(f"Processing {label}: {len(images)} images")
+        
+        for img_path in tqdm(images):
+            fname = os.path.basename(img_path)
+            save_path = os.path.join(save_label_dir, fname)
+            
+            # Execute pipeline
+            preprocess_image(img_path, save_path, 
+                             target_size=(image_size, image_size),
+                             blur_ksize=(5, 5), 
+                             adaptive_block_size=31, 
+                             adaptive_C=6,
+                             gradient_ksize=3,
+                             closing_ksize=5,
+                             min_area=100)
+
+if __name__ == "__main__":
+    main()