カルピスのモデル

app.py

@@ -36,6 +36,7 @@ if __name__ == "__main__":
                 choices=[
                     "asahachi_black",
                     "fukazara_scoop_black_4",
+                    "calpis_bottle_cap",
                 ],
                 label="Model",
             ),
@@ -48,7 +49,7 @@ if __name__ == "__main__":
                 type="text",
                 label="Threshold of difference between max and median",
                 value="0.095",
-            ),            
+            ),
             Textbox(
                 type="text",
                 label="Threshold of median",
@@ -63,11 +64,11 @@ if __name__ == "__main__":
             Textbox(
                 type="text",
                 label="Anomaly Map Max",
-            ),                       
+            ),
             Textbox(
                 type="text",
                 label="Anomaly Map Median",
-            ),            
+            ),
             Textbox(
                 type="text",
                 label="Anomaly Map",
@@ -87,4 +88,4 @@ if __name__ == "__main__":
         ],
         title="Anomaly Detection with Anomalib",
     )
-    interface.launch()
+    interface.launch()

configs/calpis_bottle_cap.yaml

@@ -0,0 +1,112 @@
+# https://github.com/openvinotoolkit/anomalib/blob/main/src/anomalib/models/patchcore/config.yaml
+
+dataset:
+  name: original
+  format: mvtec
+  path: /workspaces/anomaly-detection/datasets/original
+  task: segmentation
+  category: cap
+  train_batch_size: 32
+  eval_batch_size: 32
+  num_workers: 8
+  image_size: 256 # dimensions to which images are resized (mandatory)
+  # center_crop: 224 # dimensions to which images are center-cropped after resizing (optional)
+  normalization: imagenet # data distribution to which the images will be normalized: [none, imagenet]
+  transform_config:
+    train: null
+    eval: null
+  test_split_mode: synthetic # options: [from_dir, synthetic]
+  test_split_ratio: 0.2 # fraction of train images held out testing (usage depends on test_split_mode)
+  val_split_mode: synthetic # options: [same_as_test, from_test, synthetic]
+  val_split_ratio: 0.5 # fraction of train/test images held out for validation (usage depends on val_split_mode)
+  tiling:
+    apply: false
+    tile_size: null
+    stride: null
+    remove_border_count: 0
+    use_random_tiling: False
+    random_tile_count: 16
+
+model:
+  name: patchcore
+  backbone: wide_resnet50_2
+  pre_trained: true
+  layers:
+    - layer2
+    - layer3
+  coreset_sampling_ratio: 0.001
+  num_neighbors: 9
+  normalization_method: min_max # options: [null, min_max, cdf]
+
+metrics:
+  image:
+    - F1Score
+    - AUROC
+  pixel:
+    - F1Score
+    - AUROC
+  threshold:
+    method: adaptive #options: [adaptive, manual]
+    manual_image: null
+    manual_pixel: null
+
+visualization:
+  show_images: False # show images on the screen
+  save_images: True # save images to the file system
+  log_images: True # log images to the available loggers (if any)
+  image_save_path: null # path to which images will be saved
+  mode: full # options: ["full", "simple"]
+
+project:
+  seed: 0
+  path: ./results
+
+logging:
+  logger: [] # options: [comet, tensorboard, wandb, csv] or combinations.
+  log_graph: false # Logs the model graph to respective logger.
+
+optimization:
+  export_mode: null # options: onnx, openvino
+
+# PL Trainer Args. Don't add extra parameter here.
+trainer:
+  enable_checkpointing: true
+  default_root_dir: null
+  gradient_clip_val: 0
+  gradient_clip_algorithm: norm
+  num_nodes: 1
+  devices: 1
+  enable_progress_bar: true
+  overfit_batches: 0.0
+  track_grad_norm: -1
+  check_val_every_n_epoch: 1 # Don't validate before extracting features.
+  fast_dev_run: false
+  accumulate_grad_batches: 4
+  max_epochs: 1
+  min_epochs: null
+  max_steps: -1
+  min_steps: null
+  max_time: null
+  limit_train_batches: 0.5
+  limit_val_batches: 1.0
+  limit_test_batches: 1.0
+  limit_predict_batches: 1.0
+  val_check_interval: 1.0 # Don't validate before extracting features.
+  log_every_n_steps: 50
+  accelerator: auto # <"cpu", "gpu", "tpu", "ipu", "hpu", "auto">
+  strategy: null
+  sync_batchnorm: false
+  precision: 32
+  enable_model_summary: true
+  num_sanity_val_steps: 0
+  profiler: null
+  benchmark: false
+  deterministic: false
+  reload_dataloaders_every_n_epochs: 0
+  auto_lr_find: false
+  replace_sampler_ddp: true
+  detect_anomaly: false
+  auto_scale_batch_size: false
+  plugins: null
+  move_metrics_to_cpu: false
+  multiple_trainloader_mode: max_size_cycle

models/calpis_bottle_cap.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:755722860af37b81aa994fe94e6e734355357e8d445be2b1797b27ee2a8f34c0
+size 110947742

utils/preprocess.py

@@ -15,32 +15,61 @@ def convert_to_png(image_path):
     return png_path
 
 
-# Function to fill background and crop the image to make the object fill the frame
+OUTPUT_SIZE = 256
+EDGE_THRESHOLD = 30
+
+
 def fill_background_and_crop(image_path):
     # 背景色の設定（BGR）
-    color = [0, 0, 255]
+    color = [255, 0, 0]
 
+    # 画像を読み込み、HSV色空間に変換
     image = cv2.imread(str(image_path))
-    image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
-    gray = cv2.cvtColor(image_rgb, cv2.COLOR_RGB2GRAY)
-    blurred = cv2.GaussianBlur(gray, (5, 5), 0)
-    edges = cv2.Canny(blurred, threshold1=30, threshold2=100)
+    hsv_image = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)
+    v_channel = hsv_image[:, :, 2]
+
+    # Vチャネルに対してガウシアンブラーを適用
+    blurred_v = cv2.GaussianBlur(v_channel, (5, 5), 0)
+
+    # cv2.Cannyを使用してエッジ検出
+    edges = cv2.Canny(blurred_v, threshold1=EDGE_THRESHOLD, threshold2=100)
+
+    # エッジを太くするためにダイレーションを適用
+    kernel = np.ones((5, 5), np.uint8)
+    edges_dilated = cv2.dilate(edges, kernel, iterations=1)
+
+    # 輪郭を検出
     contours, _ = cv2.findContours(
-        edges.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+        edges_dilated, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+    if not contours:
+        print("No contours found, image unchanged.")
+        return
+
     largest_contour = max(contours, key=cv2.contourArea)
-    mask = np.zeros_like(gray)
+    mask = np.zeros_like(v_channel)
     cv2.drawContours(mask, [largest_contour], -1,
-                     color=(255,) * 3, thickness=cv2.FILLED)
+                     color=(255,), thickness=cv2.FILLED)
     mask_rgb = cv2.cvtColor(mask, cv2.COLOR_GRAY2RGB)
-    background_color = np.zeros_like(image_rgb, dtype=np.uint8)
+    background_color = np.zeros_like(image, dtype=np.uint8)
     background_color[:, :] = color
-    object_with_bg = np.where(mask_rgb == 255, image_rgb, background_color)
-    final_image = cv2.cvtColor(object_with_bg, cv2.COLOR_RGB2BGR)
-    
+    object_with_bg = np.where(mask_rgb == 255, image, background_color)
+
     # Find the bounding rect and crop the image
     x, y, w, h = cv2.boundingRect(largest_contour)
-    cropped_image = final_image[y:y+h, x:x+w]
-    cropped_image = cv2.resize(cropped_image, (256, 256), interpolation=cv2.INTER_AREA)
-    
+    cropped_image = object_with_bg[y:y+h, x:x+w]
+    cropped_image = cv2.GaussianBlur(cropped_image, (5, 5), 0)
+
+    # リサイズ処理の最適化
+    cropped_image = resize_image_optimized(cropped_image, OUTPUT_SIZE)
+
     # Overwrite the original image
     cv2.imwrite(str(image_path), cropped_image)
+
+
+def resize_image_optimized(image, size):
+    h, w = image.shape[:2]
+    ratio = size / max(h, w)
+    new_h, new_w = int(h * ratio), int(w * ratio)
+    resized_image = cv2.resize(
+        image, (new_w, new_h), interpolation=cv2.INTER_LINEAR)
+    return resized_image