{"id":5672,"date":"2025-07-09T10:48:44","date_gmt":"2025-07-09T01:48:44","guid":{"rendered":"https:\/\/www.wetch.co.jp\/note\/?p=5672"},"modified":"2025-07-09T10:48:45","modified_gmt":"2025-07-09T01:48:45","slug":"%e3%83%a9%e3%83%99%e3%83%ab%e7%94%bb%e5%83%8f%e3%81%ae%e5%88%87%e3%82%8a%e5%87%ba%e3%81%97","status":"publish","type":"post","link":"https:\/\/www.wetch.co.jp\/note\/%e3%83%a9%e3%83%99%e3%83%ab%e7%94%bb%e5%83%8f%e3%81%ae%e5%88%87%e3%82%8a%e5%87%ba%e3%81%97\/","title":{"rendered":"\u30e9\u30d9\u30eb\u753b\u50cf\u306e\u5207\u308a\u51fa\u3057"},"content":{"rendered":"\n

\u524d\u56de\u306e\u6295\u7a3f\uff08Mac+PostgreSQL\u3067\u30d9\u30af\u30c8\u30eb\u691c\u7d22<\/a>\uff09\u3067\u3001\u300c\u6b21\u56de\u306f\u3001\u5148\u306b\u751f\u6210\u3057\u305f\u30e2\u30c7\u30eb\u3092\u4f7f\u7528\u3057\u3066\u30e9\u30d9\u30eb\u90e8\u5206\u3092\u30c8\u30ea\u30df\u30f3\u30b0\u3057\u3001\u30d9\u30af\u30c8\u30eb\u5316\u3057\u305f\u4e0a\u3067\u306e\u691c\u7d22\u7cbe\u5ea6\u306b\u95a2\u3057\u3066\u78ba\u8a8d\u3057\u3066\u307f\u308b\u300d\u3068\u98a8\u5442\u6577\u3092\u5e83\u3052\u307e\u3057\u305f\u304c\u3001\u3044\u3056\u4f5c\u696d\u3092\u59cb\u3081\u3066\u307f\u308b\u3068\u3001\u30e9\u30d9\u30eb\u90e8\u5206\u3092\u5207\u308a\u51fa\u3059\u3060\u3051\u3067\u7d50\u69cb\u306a\u4f5c\u696d\u304c\u767a\u751f\u3057\u307e\u3057\u305f\u306e\u3067\u3001\u4e00\u65e6\u305d\u3053\u307e\u3067\u306e\u5185\u5bb9\u3092\u307e\u3068\u3081\u305f\u3044\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\u306a\u304a\u3001\u6628\u4eca\u306e\u95a2\u9023\u6295\u7a3f\u306b\u304a\u3044\u3066\u306f\u300c\u30e9\u30d9\u30eb\u753b\u50cf\u306e\u30c8\u30ea\u30df\u30f3\u30b0\u300d\u3068\u9298\u6253\u3063\u3066\u30b7\u30ea\u30fc\u30ba\u5316\u3057\u3066\u304d\u307e\u3057\u305f\u304c\u3001\u300c\u30c8\u30ea\u30df\u30f3\u30b0\u300d\u3068\u306f\u300c\u753b\u50cf\u5168\u4f53\u306e\u69cb\u56f3\u3092\u6574\u3048\u308b\u305f\u3081\u306b\u3001\u4e0d\u8981\u306a\u90e8\u5206\u3092\u5207\u308a\u843d\u3068\u3059<\/strong>\u3053\u3068\u3092\u6307\u3059\u306e\u304c\u4e00\u822c\u7684\u300d\uff08Gemini\u5148\u751f\u8ac7\uff09\u3068\u306e\u3053\u3068\u3067\u3001\u30e9\u30d9\u30eb\u90e8\u5206\u3092\u5207\u308a\u51fa\u3059\u3068\u3044\u3046\u672c\u4f5c\u696d\u306e\u8da3\u65e8\u3068\u306f\u82e5\u5e72\u7570\u306a\u308b\u610f\u5473\u306b\u306a\u308b\u3088\u3046\u306a\u306e\u3067\u3001\u4eca\u5f8c\u306f\u30c8\u30ea\u30df\u30f3\u30b0\u3068\u3044\u3046\u8868\u73fe\u306f\u4f7f\u308f\u306a\u3044\u3053\u3068\u306b\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

\u95a2\u9023\u30d1\u30c3\u30b1\u30fc\u30b8\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb<\/h2>\n\n\n\n

\u65b0\u305f\u306bPython\u306evenv\u74b0\u5883\u3092\u69cb\u7bc9\u3057\u3001\u5fc5\u8981\u306a\u30d1\u30c3\u30b1\u30fc\u30b8\u3092\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3057\u307e\u3059\u3002
\u73fe\u6642\u70b9\u306e\u4f5c\u696d\u306b\u5fc5\u8981\u306a\u30d1\u30c3\u30b1\u30fc\u30b8\u306f\u4ee5\u4e0b\u306e3\u3064\u3067\u3059\u3002<\/p>\n\n\n\n

# pip install tensorflow tensorflow-metal opencv-python<\/code><\/pre>\n\n\n\n
\u30e2\u30c7\u30eb\u306e\u30ed\u30fc\u30c9<\/h2>\n\n\n\n
\u5f53\u7136\u306a\u304c\u3089\u3001\u30e9\u30d9\u30eb\u90e8\u5206\u306e\u7279\u5b9a\u306b\u306f\u4eca\u307e\u3067\u82e6\u52b4\u3057\u3066\u80b2\u3066\u3066\u304d\u305f\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u7528\u30e2\u30c7\u30eb\u3092\u4f7f\u7528\u3057\u307e\u3059\u3002
\u5b66\u7fd2\u6642\u306b\u3082\u65e2\u5b58\u30e2\u30c7\u30eb\u306e\u8aad\u307f\u8fbc\u307f\u51e6\u7406\u306f\u3042\u308a\u307e\u3057\u305f\u304c\u3001\u6539\u3081\u3066\u4eca\u56de\u306f\u4ee5\u4e0b\u306e\u3088\u3046\u306a\u95a2\u6570\u3092\u7528\u610f\u3057\u307e\u3057\u305f\u3002<\/p>\n\n\n\n
import tensorflow as tf\n\nfrom train_model import dice_loss\nfrom train_model import dice_coef\n\ndef get_model(model_path):\n    try:\n        model = tf.keras.models.load_model(\n            model_path,\n            custom_objects={'dice_loss': dice_loss, 'dice_coef': dice_coef}\n        )\n        return model\n    except Exception as e:\n        print(f\"Error loading model: {e}\")\n        exit()<\/code><\/pre>\n\n\n\n
\u640d\u5931\u95a2\u6570\u3001\u8a55\u4fa1\u6307\u6a19\u95a2\u6570\u306f\u5b66\u7fd2\u6642\u306b\u5229\u7528\u3057\u305ftrain_model.py\u5185\u306e\u95a2\u6570\u3092\u6307\u5b9a\u3057\u307e\u3059\u3002
\u3053\u308c\u3089\u95a2\u6570\u306f\u3042\u304f\u307e\u3067\u5b66\u7fd2\u7528\u306b\u5fc5\u8981\u306a\u3082\u306e\u3067\u3001\u4eca\u56de\u306e\u3088\u3046\u306b\u5f53\u8a72\u30e2\u30c7\u30eb\u3067\u4e88\u6e2c\u3092\u5b9f\u65bd\u3059\u308b\u6bb5\u968e\u3067\u306f\u4f7f\u308f\u306a\u3044\u306e\u3067\u3059\u304c\u3001\u540c\u95a2\u6570\u3092\u4f7f\u3063\u3066\u5b66\u7fd2\u3092\u884c\u306a\u3063\u305f\u30e2\u30c7\u30eb\u306e\u30ed\u30fc\u30c9\u306b\u969b\u3057\u3066\u306f\u3001\u305d\u308c\u3089\u95a2\u6570\u3078\u306e\u53c2\u7167\u304c\u6709\u52b9\u306b\u306a\u3063\u3066\u3044\u308b\u5fc5\u8981\u304c\u3042\u308b\u3088\u3046\u3067\u3059\u3002<\/p>\n\n\n\n
\u30de\u30b9\u30af\u753b\u50cf\u306e\u751f\u6210<\/h2>\n\n\n\n
\u30de\u30b9\u30af\u753b\u50cf\u306e\u751f\u6210\u306b\u95a2\u3057\u3066\u306f\u4ee5\u4e0b\u306e\u901a\u308a\u3067\u3059\u3002<\/p>\n\n\n\n
import cv2\nimport numpy as np\nimport tensorflow as tf\nimport os\n\nIMAGE_SIZE = (256, 256)\n\ndef preprocess_image(image_path, image_size):\n    image = cv2.imread(image_path)\n    image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n    image = cv2.resize(image, image_size)\n    image = image \/ 255.0\n    return image.astype(np.float32)\n\ndef infer_mask(model, image):\n    image = np.expand_dims(image, axis=0)\n    prediction = model.predict(image)\n    mask = prediction.squeeze()\n    return mask\n\ndef extract_largest_mask(mask_image: np.ndarray, binary_threshold: int = 127) -> np.ndarray:\n    temp_mask = (np.clip(mask_image, 0, 1) * 255).astype(np.uint8)\n    _, binary_mask = cv2.threshold(temp_mask, binary_threshold, 255, cv2.THRESH_BINARY)\n\n    contours, _ = cv2.findContours(binary_mask.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)\n\n    max_area = 0\n    max_contour = None\n    for contour in contours:\n        area = cv2.contourArea(contour)\n        if area > max_area:\n            max_area = area\n            max_contour = contour\n\n    largest_object_mask = np.zeros(binary_mask.shape, dtype=np.uint8)\n    if max_contour is not None:\n        cv2.drawContours(largest_object_mask, [max_contour], -1, 255, cv2.FILLED)\n    \n    return largest_object_mask\n\ndef get_mask(model, image_path,\n             binary_threshold: int = 127):\n    image = preprocess_image(image_path, IMAGE_SIZE)\n\n    raw_mask_prediction = infer_mask(model, image)\n\n    final_mask = extract_largest_mask(raw_mask_prediction, binary_threshold)\n\n    return final_mask<\/code><\/pre>\n\n\n\n
\u5168\u4f53\u7684\u306b\u306f\u4eca\u307e\u3067\u3082\u5b9f\u65bd\u3057\u3066\u304d\u305f\u51e6\u7406\u3067\u3059\u304c\u3001\u7279\u7b46\u3059\u3079\u304d\u306f\u4eca\u56de\u8ffd\u52a0\u3057\u305f\u95a2\u6570extract_largest_mask\u306e\u5185\u5bb9\u3067\u3059\u3002<\/p>\n\n\n\n
\u4eca\u307e\u3067\u306e\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u306b\u304a\u3044\u3066\u3001\u751f\u6210\u3055\u308c\u305f\u30de\u30b9\u30af\u753b\u50cf\u306b\u306f\u30e9\u30d9\u30eb\u4ee5\u5916\u306e\u90e8\u5206\u3092\u5bfe\u8c61\u3068\u8aa4\u8a8d\u3057\u305f\u7b87\u6240\u3084\u3001\u9006\u306b\u30e9\u30d9\u30eb\u5185\u306e\u4e00\u90e8\u3092\u5bfe\u8c61\u5916\u3068\u8aa4\u8a8d\u3057\u305f\u7b87\u6240\u304c\u76f8\u5f53\u6570\u898b\u53d7\u3051\u3089\u308c\u307e\u3057\u305f\u3002
extract_largest_mask\u306f\u305d\u308c\u3089\u3078\u306e\u5bfe\u51e6\u3092\u884c\u3046\u95a2\u6570\u3067\u3059\u3002
\u5177\u4f53\u7684\u306b\u306f\u3001\u8907\u6570\u306e\u30de\u30b9\u30af\u7b87\u6240\uff08\u30de\u30b9\u30af\u753b\u50cf\u5185\u306e\u767d\u304f\u5857\u3089\u308c\u305f\u7b87\u6240\uff09\u306e\u305d\u308c\u305e\u308c\u306b\u95a2\u3057\u3066\u8f2a\u90ed\u3092\u62bd\u51fa\u3057\u3001\u305d\u306e\u4e2d\u304b\u3089\u6700\u5927\u306e\u9762\u7a4d\u3092\u6301\u3064\u3082\u306e\u3092\u62bd\u51fa\u3057\u3066\u3001\u305d\u306e\u5185\u90e8\u3092\u5168\u3066\u767d\u304f\u5857\u308a\u6f70\u3057\u305f\u30de\u30b9\u30af\u753b\u50cf\u3092\u65b0\u305f\u306b\u751f\u6210\u3057\u3066\u3044\u307e\u3059\u3002
\u3053\u308c\u306b\u3088\u308a\u3001\u30e9\u30d9\u30eb\u90e8\u5206\u4ee5\u5916\u306e\u96e2\u308c\u5c0f\u5cf6\u306e\u3088\u3046\u306a\u30de\u30b9\u30af\u7b87\u6240\u306f\u65b0\u30de\u30b9\u30af\u753b\u50cf\u751f\u6210\u6642\u306b\u7121\u8996\u3055\u308c\u307e\u3059\u3057\u3001\u30e9\u30d9\u30eb\u90e8\u5206\u306b\u30b7\u30df\u306e\u3088\u3046\u306b\u751f\u3058\u305f\u9ed2\u3044\u7b87\u6240\u3082\u767d\u304f\u5857\u308a\u6f70\u3055\u308c\u308b\u306e\u3067\u3001\u672c\u6765\u306e\u30e9\u30d9\u30eb\u90e8\u5206\u3092\u5207\u308a\u629c\u304f\u305f\u3081\u306e\u30de\u30b9\u30af\u3068\u3057\u3066\u3001\u3088\u308a\u9069\u5207\u306a\u3082\u306e\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n
\u30e9\u30d9\u30eb\u753b\u50cf\u306e\u751f\u6210<\/h2>\n\n\n\n
\u5148\u306b\u751f\u6210\u3055\u308c\u305f\u30de\u30b9\u30af\u3092\u5143\u753b\u50cf\u306b\u9069\u7528\u3059\u308b\u3053\u3068\u3067\u3001\u30e9\u30d9\u30eb\uff08\u3068\u601d\u308f\u308c\u308b\uff09\u90e8\u5206\u3092\u305d\u306e\u307e\u307e\u6b8b\u3057\u3001\u305d\u308c\u4ee5\u5916\u3092\u9ed2\u304f\u5857\u308a\u6f70\u3057\u305f\u753b\u50cf\u304c\u751f\u6210\u3067\u304d\u307e\u3059\u3002
\u305f\u3060\u3001\u3053\u308c\u3060\u3051\u3067\u306f\u30e9\u30d9\u30eb\u306e\u4f4d\u7f6e\u3082\u30b5\u30a4\u30ba\u3082\u69d8\u3005\u306a\u753b\u50cf\u304c\u3067\u304d\u3066\u3057\u307e\u3046\u306e\u3067\u3001\u5207\u308a\u51fa\u3057\u305f\u30e9\u30d9\u30eb\u90e8\u5206\u304c\u6240\u5b9a\u306e\u30b5\u30a4\u30ba\uff08\u305d\u306e\u5f8c\u306e\u30d9\u30af\u30c8\u30eb\u5316\u3092\u8003\u616e\u3057\u3066224×224\u3092\u60f3\u5b9a\uff09\u3067\u5927\u5199\u3057\u3055\u308c\u305f\u753b\u50cf\u306b\u52a0\u5de5\u3057\u305f\u3044\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n\n\n\n
\u5177\u4f53\u7684\u306a\u5185\u5bb9\u306f\u4ee5\u4e0b\u306e\u901a\u308a\uff08\u4f8b\u306b\u3088\u3063\u3066Gemini\u5148\u751f\u4f5c\uff09\u3002<\/p>\n\n\n\n
import cv2\nimport numpy as np\n\ndef apply_mask_and_crop(org_image_path, mask, target_crop_size: int = 224) -> np.ndarray:\n    org_image = cv2.imread(org_image_path)\n    org_image = cv2.cvtColor(org_image, cv2.COLOR_BGR2RGB)\n    org_height, org_width, _ = org_image.shape\n\n    resized_mask = cv2.resize(mask, (org_width, org_height), interpolation=cv2.INTER_NEAREST)\n    resized_mask_expanded = np.expand_dims(resized_mask, axis=-1)\n    masked_image = cv2.bitwise_and(org_image, org_image, mask=resized_mask)\n\n    contours, _ = cv2.findContours(resized_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)\n    x, y, w, h = cv2.boundingRect(contours[0])\n    min_x, min_y = x, y\n    max_x, max_y = x + w, y + h\n    min_x, min_y = max(0, min_x), max(0, min_y)\n    max_x, max_y = min(org_width, max_x), min(org_height, max_y)\n    \n    bbox_width = max_x - min_x\n    bbox_height = max_y - min_y\n\n    crop_width_orig = max(bbox_width, bbox_height)\n    crop_height_orig = max(bbox_width, bbox_height)\n\n    center_x = min_x + bbox_width \/\/ 2\n    center_y = min_y + bbox_height \/\/ 2\n\n    crop_x1 = center_x - crop_width_orig \/\/ 2\n    crop_y1 = center_y - crop_height_orig \/\/ 2\n    crop_x2 = crop_x1 + crop_width_orig\n    crop_y2 = crop_y1 + crop_height_orig\n\n    pad_left = max(0, -crop_x1)\n    pad_top = max(0, -crop_y1)\n    pad_right = max(0, crop_x2 - org_width)\n    pad_bottom = max(0, crop_y2 - org_height)\n    padded_masked_image = cv2.copyMakeBorder(masked_image, \n                                               pad_top, pad_bottom, \n                                               pad_left, pad_right, \n                                               cv2.BORDER_CONSTANT, value=[0, 0, 0])\n\n    crop_x1_padded = crop_x1 + pad_left\n    crop_y1_padded = crop_y1 + pad_top\n    crop_x2_padded = crop_x2 + pad_left\n    crop_y2_padded = crop_y2 + pad_top\n    cropped_image = padded_masked_image[crop_y1_padded:crop_y2_padded, crop_x1_padded:crop_x2_padded]\n    \n    final_cropped_image = cv2.resize(cropped_image,\n                                     (target_crop_size, target_crop_size),\n                                     interpolation=cv2.INTER_AREA)\n    \n    return final_cropped_image<\/code><\/pre>\n\n\n\n
\u30dd\u30a4\u30f3\u30c8\u306b\u306a\u308a\u305d\u3046\u306a\u90e8\u5206\u3092\u500b\u5225\u306b\u898b\u3066\u3044\u304d\u307e\u3059\u3002<\/p>\n\n\n\n
    resized_mask = cv2.resize(mask, (org_width, org_height), interpolation=cv2.INTER_NEAREST)\n    resized_mask_expanded = np.expand_dims(resized_mask, axis=-1)\n    masked_image = cv2.bitwise_and(org_image, org_image, mask=resized_mask)<\/code><\/pre>\n\n\n\n
\u30e2\u30c7\u30eb\u306e\u5b66\u7fd2\u306b\u969b\u3057\u3066\u306f\u5143\u753b\u50cf\uff08org_image\uff09\u3092256×256\u306b\u30ea\u30b5\u30a4\u30ba\u3057\u3066\u3044\u305f\u306e\u3067\u3001\u540c\u30e2\u30c7\u30eb\u304c\u751f\u6210\u3057\u305f\u30de\u30b9\u30af\u753b\u50cf\u3082\u540c\u3058\u30b5\u30a4\u30ba\uff08256×256\uff09\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u3002
\u3053\u306e\u30de\u30b9\u30af\u3092\u30ea\u30b5\u30a4\u30ba\u5f8c\u306e\u753b\u50cf\u306b\u9069\u7528\u3059\u308b\u3068\u3001\u30e9\u30d9\u30eb\u90e8\u5206\u306f\u304b\u306a\u308a\u5c0f\u3055\u3044\uff08\u60c5\u5831\u91cf\u306e\u5c11\u306a\u3044\uff09\u3082\u306e\u306b\u306a\u3063\u3066\u3057\u307e\u3044\u307e\u3059\u3002
\u3088\u3063\u3066\u3001\u30ea\u30b5\u30a4\u30ba\u524d\u306e\u5143\u753b\u50cf\u306b\u5bfe\u3057\u3066\u30de\u30b9\u30af\u304c\u9069\u7528\u3067\u304d\u308b\u3088\u3046\u3001\u30de\u30b9\u30af\u753b\u50cf\u306e\u65b9\u3092\u5143\u753b\u50cf\u306b\u5408\u308f\u305b\u3066\u62e1\u5927\u3057\u3001\u305d\u306e\u7d50\u679c\u3092\u5143\u753b\u50cf\u306b\u9069\u7528\u3092\u3057\u3066\u3044\u307e\u3059\u3002
\u3053\u308c\u306b\u3088\u308a\u3001\u30b5\u30a4\u30ba\u306f\u5143\u753b\u50cf\u3068\u540c\u3058\u3067\u3001\u30e9\u30d9\u30eb\u90e8\u5206\u4ee5\u5916\u304c\u9ed2\u304f\u5857\u308a\u3064\u3076\u3055\u308c\u305f\u753b\u50cf\uff08masked_image\uff09\u304c\u751f\u6210\u3055\u308c\u307e\u3059\u3002<\/p>\n\n\n\n
    contours, _ = cv2.findContours(resized_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)\n    x, y, w, h = cv2.boundingRect(contours[0])\n    min_x, min_y = x, y\n    max_x, max_y = x + w, y + h\n    min_x, min_y = max(0, min_x), max(0, min_y)\n    max_x, max_y = min(org_width, max_x), min(org_height, max_y)\n    \n    bbox_width = max_x - min_x\n    bbox_height = max_y - min_y\n\n    crop_width_orig = max(bbox_width, bbox_height)\n    crop_height_orig = max(bbox_width, bbox_height)<\/code><\/pre>\n\n\n\n
masked_image\u306b\u306f\u30e9\u30d9\u30eb\u4ee5\u5916\u306e\u4f59\u8a08\u306a\u90e8\u5206\u304c\u5927\u91cf\u306b\u542b\u307e\u308c\u307e\u3059\u306e\u3067\u3001\u305d\u308c\u3089\u3092\u6975\u529b\u9664\u5916\u3057\u3001\u30e9\u30d9\u30eb\u90e8\u5206\u304c\u30ae\u30ea\u30ae\u30ea\u7d0d\u307e\u308b\u6b63\u65b9\u5f62\u306e\u753b\u50cf\u3068\u3057\u3066\u5207\u308a\u51fa\u305b\u308b\u3088\u3046\u3001\u4f4d\u7f6e\u3084\u30b5\u30a4\u30ba\u3092\u8a08\u7b97\u3057\u307e\u3059\u3002
\u307e\u305a\u3001cv2.findContours\u3067\u30de\u30b9\u30af\u753b\u50cf\u304b\u3089\u30de\u30b9\u30af\u306e\u8f2a\u90ed\u3092\u62bd\u51fa\u3057\u3001cv2.boundingRect\u3067\u305d\u308c\u3092\u30ae\u30ea\u30ae\u30ea\u56f2\u3080\u77e9\u5f62\u306e\u5ea7\u6a19\u3068\u30b5\u30a4\u30ba\u3092\u7b97\u51fa\u3057\u3066\u3044\u307e\u3059\u3002
bbox_width, bbox_height\u306f\u524d\u8ff0\u306e\u77e9\u5f62\u306e\u5e45\u3068\u9ad8\u3055\u306b\u306a\u308a\u3001crop_width_orig, crop_height_orig\u306f\u305d\u308c\u3092\u5305\u542b\u3059\u308b\u6b63\u65b9\u5f62\u306e\u5e45\u3068\u9ad8\u3055\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n
\u306a\u304a\u3001\u82e5\u5e72\u5197\u9577\u306a\u51e6\u7406\u3092\u884c\u306a\u3063\u3066\u3044\u308b\u3088\u3046\u306b\u898b\u3048\u308b\u90e8\u5206\u3082\u3042\u308a\u307e\u3059\u304c\u3001Gemini\u5148\u751f\u306a\u308a\u306e\u3053\u3060\u308f\u308a\u304c\u3042\u308b\u3088\u3046\u3067\u3059\uff08\u5358\u306b\u8a08\u7b97\u7684\u306b\u5408\u3063\u3066\u3044\u308b\u304b\u3069\u3046\u304b\u3067\u306f\u306a\u304f\u3001\u300c\u30b3\u30fc\u30c9\u3092\u66f8\u304f\u4eba\u306e”\u610f\u56f3”\u3068”\u601d\u8003\u306e\u30d7\u30ed\u30bb\u30b9”\u3092\u793a\u3059\u305f\u3081\u3001\u3042\u308b\u3044\u306f\u6c4e\u7528\u7684\u306a\u30d1\u30bf\u30fc\u30f3\u306e\u9069\u7528\u306e\u305f\u3081\u300d\u3068\u306e\u3053\u3068\uff09\u3002<\/p>\n\n\n\n
    center_x = min_x + bbox_width \/\/ 2\n    center_y = min_y + bbox_height \/\/ 2\n\n    crop_x1 = center_x - crop_width_orig \/\/ 2\n    crop_y1 = center_y - crop_height_orig \/\/ 2\n    crop_x2 = crop_x1 + crop_width_orig\n    crop_y2 = crop_y1 + crop_height_orig<\/code><\/pre>\n\n\n\n
\u5148\u306e\u8a08\u7b97\u7d50\u679c\u3092\u5143\u306b\u3001\u30e9\u30d9\u30eb\u90e8\u5206\u304c\u4e2d\u5fc3\u306b\u4f4d\u7f6e\u3059\u308b\u3088\u3046\u306b\u8003\u616e\u3057\u3064\u3064\u3001\u5207\u308a\u51fa\u3059\u6b63\u65b9\u5f62\u306e\u5ea7\u6a19\u3092\u7b97\u51fa\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n
    pad_left = max(0, -crop_x1)\n    pad_top = max(0, -crop_y1)\n    pad_right = max(0, crop_x2 - org_width)\n    pad_bottom = max(0, crop_y2 - org_height)\n    padded_masked_image = cv2.copyMakeBorder(masked_image, \n                                               pad_top, pad_bottom, \n                                               pad_left, pad_right, \n                                               cv2.BORDER_CONSTANT, value=[0, 0, 0])<\/code><\/pre>\n\n\n\n
\u5143\u753b\u50cf\u304b\u3089\u30e9\u30d9\u30eb\u90e8\u5206\u3092\u542b\u3080\u6b63\u65b9\u5f62\u3092\u5207\u308a\u51fa\u3059\u5834\u5408\u3001\u30e9\u30d9\u30eb\u90e8\u5206\u304c\u5143\u753b\u50cf\u306e\u7aef\u306e\u65b9\u306b\u4f4d\u7f6e\u3057\u3066\u3044\u305f\u5834\u5408\u306a\u3069\u306f\u5207\u308a\u51fa\u3059\u6b63\u65b9\u5f62\u304c\u5143\u753b\u50cf\u5185\u306b\u7d0d\u307e\u3089\u306a\u3044\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002
\u3088\u3063\u3066\u3001\u6b63\u65b9\u5f62\u304c\u5143\u753b\u50cf\u304b\u3089\u306f\u307f\u51fa\u3059\u5834\u5408\u306f\u3001\u305d\u306e\u5206\u5143\u753b\u50cf\u3092\u62e1\u5f35\u3057\u3001\u8ffd\u52a0\u3057\u305f\u90e8\u5206\u306f\u9ed2\u304f\u5857\u308a\u3064\u3076\u3059\u3068\u3044\u3046\u51e6\u7406\u3092\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n
\u86c7\u8db3\u306a\u304c\u3089\u3001\u4e0a\u8a18\u53ef\u80fd\u6027\u306b\u3064\u3044\u3066\u306f\u3001\u500b\u4eba\u7684\u306b\u306f\u76f2\u70b9\u3067\u3057\u305f\u3002
\u3055\u3059\u304cGemini\u5148\u751f\u3002<\/p>\n\n\n\n
    crop_x1_padded = crop_x1 + pad_left\n    crop_y1_padded = crop_y1 + pad_top\n    crop_x2_padded = crop_x2 + pad_left\n    crop_y2_padded = crop_y2 + pad_top\n    cropped_image = padded_masked_image[crop_y1_padded:crop_y2_padded, crop_x1_padded:crop_x2_padded]<\/code><\/pre>\n\n\n\n
\u5148\u306e\u51e6\u7406\u3067\u5143\u753b\u50cf\u306e\u5f62\u304c\u5909\u308f\u3063\u3066\u3044\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u3001\u305d\u306e\u5834\u5408\u306f\u5207\u308a\u629c\u304f\u7bc4\u56f2\u306e\u5ea7\u6a19\u3082\u5909\u308f\u3063\u3066\u3044\u308b\u306e\u3067\u3001\u305d\u306e\u70b9\u3092\u8003\u616e\u3057\u3066\u5ea7\u6a19\u3092\u88dc\u6b63\u3057\u3064\u3064\u3001\u30e9\u30d9\u30eb\u90e8\u5206\uff08cropped_image\uff09\u3092\u5207\u308a\u629c\u304d\u307e\u3059\u3002<\/p>\n\n\n\n
    final_cropped_image = cv2.resize(cropped_image,\n                                     (target_crop_size, target_crop_size),\n                                     interpolation=cv2.INTER_AREA)<\/code><\/pre>\n\n\n\n
\u5207\u308a\u629c\u3044\u305f\u30e9\u30d9\u30eb\u90e8\u5206\uff08\u6b63\u65b9\u5f62\uff09\u306e\u30b5\u30a4\u30ba\u306f\u69d8\u3005\u306a\u306e\u3067\u3001\u3053\u308c\u3092\u6240\u5b9a\u306e\u30b5\u30a4\u30ba\uff08\u30c7\u30d5\u30a9\u30eb\u30c8\u306f224×224\uff09\u306b\u30ea\u30b5\u30a4\u30ba\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
\u4e0a\u8a18\u51e6\u7406\u306b\u3088\u308a\u3001\u5143\u753b\u50cf\uff08org_image\uff09\u306b\u304a\u3044\u3066\u30e9\u30d9\u30eb\u90e8\u5206\u304c\u3069\u306e\u3088\u3046\u306a\u30b5\u30a4\u30ba\u3067\u3069\u306e\u3088\u3046\u306a\u4f4d\u7f6e\u306b\u5199\u3063\u3066\u3044\u3066\u3082\u3001\u30e9\u30d9\u30eb\u90e8\u5206\u304c\u4e2d\u592e\u306b\u3001\u304b\u3064\u6700\u5927\u30b5\u30a4\u30ba\u3067\u5199\u3055\u308c\u305f\u540c\u3058\u30b5\u30a4\u30ba\u306e\u753b\u50cf\uff08final_cropped_image\uff09\u3068\u3057\u3066\u52a0\u5de5\u3055\u308c\u308b\u3053\u3068\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n
\u30e9\u30d9\u30eb\u753b\u50cf\u306e\u4fdd\u5b58<\/h2>\n\n\n\n
\u524d\u8ff0\u3057\u305f\u51e6\u7406\u3092\u5b9f\u884c\u3057\u3064\u3064\u3001\u751f\u6210\u3055\u308c\u305f\u30e9\u30d9\u30eb\u753b\u50cf\u3092\u6240\u5b9a\u306e\u30c7\u30a3\u30ec\u30af\u30c8\u30ea\u914d\u4e0b\u306b\u4fdd\u5b58\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
import os\nimport cv2\n\nfrom get_model import get_model\nfrom get_mask import get_mask\nfrom apply_mask_and_crop import apply_mask_and_crop\n\nMODEL_PATH = \".\/my_model\/unet.keras\"\nINPUTS = \".\/inputs\"\nOUTPUTS = \".\/outputs\"\nCROP_SIZE = 224\n\ndef save_img(img, save_path):\n    if img.ndim == 3 and img.shape[2] == 3:\n        img_to_save = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)\n    else:\n        img_to_save = img\n\n    cv2.imwrite(save_path, img_to_save)\n\n# MAIN\nos.makedirs(OUTPUTS, exist_ok=True)\n\nmodel = get_model(MODEL_PATH)\n\nimage_files = os.listdir(INPUTS)\nfor image_file in image_files:\n    print(f\"Processing {image_file}...\")\n    image_path = os.path.join(INPUTS, image_file)\n\n    mask = get_mask(model, image_path)\n\n    save_path = os.path.join(OUTPUTS, os.path.splitext(image_file)[0] + '_mask.png')\n    save_img(mask, save_path)\n\n    cropped_masked_image = apply_mask_and_crop(image_path, mask, CROP_SIZE)\n\n    base_name, original_ext = os.path.splitext(image_file)\n    save_path_cropped = os.path.join(OUTPUTS, base_name + '_cropped' + original_ext)\n    save_img(cropped_masked_image, save_path_cropped)<\/code><\/pre>\n\n\n\n
\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u306b\u4f7f\u7528\u3059\u308b\u30e2\u30c7\u30eb\u306f\u300c.\/my_model\/unet.keras\u300d\u3068\u3057\u3066\u914d\u7f6e\u3057\u3066\u304a\u304d\u307e\u3059\u3002
\u4eca\u56de\u306f\u300c\u30e9\u30d9\u30eb\u753b\u50cf\u306e\u30c8\u30ea\u30df\u30f3\u30b0\uff08\uff16\uff09<\/a>\u300d\u3067val_loss\u306e\u5024\u304c\u6700\u3082\u5c0f\u3055\u304b\u3063\u305f\uff08\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u7d50\u679c\u3082\u7dcf\u5408\u7684\u306b\u826f\u3055\u305d\u3046\u306a\uff09\u300c\u5b66\u7fd2\u7387\u6e1b\u8870\u300d\u30e2\u30c7\u30eb\u3092\u4f7f\u7528\u3059\u308b\u3053\u3068\u306b\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
\u5b9f\u884c\u7d50\u679c<\/h2>\n\n\n\n
\u30e9\u30d9\u30eb\u753b\u50cf\u306e\u5207\u308a\u51fa\u3057\u7d50\u679c\u3092\u4ee5\u4e0b\u306b\u793a\u3057\u307e\u3059\u3002
\u4e0a\u304c\u5143\u753b\u50cf\u3067\u4e0b\u304c\u5207\u308a\u51fa\u3057\u7d50\u679c\u3067\u3059\u3002<\/p>\n\n\n\n
\u30bb\u30c3\u30c8\uff11<\/h3>\n\n\n\n
\"\"<\/td>\"\"<\/td>\"\"<\/td>\"\"<\/td>\"\"<\/td><\/tr>
\"\"<\/td>\"\"<\/td>\"\"<\/td>\"\"<\/td>\"\"<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
\u30bb\u30c3\u30c8\uff12<\/h3>\n\n\n\n
\"\"<\/td>\"\"<\/td>\"\"<\/td>\"\"<\/td>\"\"<\/td><\/tr>
\"\"<\/td>\"\"<\/td>\"\"<\/td>\"\"<\/td>\"\"<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
\u30bb\u30c3\u30c8\uff13<\/h3>\n\n\n\n
\"\"<\/td>\"\"<\/td>\"\"<\/td>\"\"<\/td>\"\"<\/td><\/tr>
\"\"<\/td>\"\"<\/td>\"\"<\/td>\"\"<\/td>\"\"<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
\u30bb\u30c3\u30c8\uff14<\/h3>\n\n\n\n
\"\"<\/td>\"\"<\/td>\"\"<\/td>\"\"<\/td>\"\"<\/td><\/tr>
\"\"<\/td>\"\"<\/td>\"\"<\/td>\"\"<\/td>\"\"<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
\u307e\u3068\u3081<\/h2>\n\n\n\n
\u7387\u76f4\u306b\u8a00\u3063\u3066\u3001\u304b\u306a\u308a\u826f\u3044\u306e\u3067\u306f\u306a\u3044\u3067\u3057\u3087\u3046\u304b\uff01
\u4e00\u90e8\u30e9\u30d9\u30eb\u4ee5\u5916\u306e\u90e8\u5206\u3092\u542b\u3093\u3067\u3057\u307e\u3063\u3066\u3044\u308b\u30b1\u30fc\u30b9\u3082\u3042\u308a\u307e\u3059\u304c\u3001\u30d9\u30af\u30c8\u30eb\u5316\u3092\u60f3\u5b9a\u3057\u3066\u753b\u50cf\u5185\u306e\u30e9\u30d9\u30eb\u90e8\u5206\u3092\u540c\u7b49\u306e\u30b5\u30a4\u30ba\u30fb\u69cb\u56f3\u3067\u753b\u50cf\u5316\u3059\u308b\u3068\u8a00\u3046\u76ee\u7684\u306b\u7167\u3089\u305b\u3070\u3001\u5341\u5206\u6e80\u8db3\u3067\u304d\u308b\u30ec\u30d9\u30eb\u3067\u306f\u306a\u3044\u304b\u3068\u601d\u3063\u3066\u3044\u307e\u3059\u3002
\u7279\u306b\u30bb\u30c3\u30c8\uff14\u306a\u3069\u306f\u30d9\u30b9\u30c8\u306b\u8fd1\u3044\u7d50\u679c\u304b\u3068\u81ea\u8ca0\u3057\u3066\u307e\u3059\u3002<\/p>\n\n\n\n
\u304b\u3064\u3066\u300c\u753b\u50cf\u691c\u7d22\u306e\u7cbe\u5ea6\u78ba\u8a8d\uff08\uff15\uff09<\/a>\u300d\u306b\u304a\u3044\u3066rembg\u3092\u4f7f\u7528\u3057\u305f\u30e9\u30d9\u30eb\u306e\u5207\u308a\u629c\u304d\u306b\u632b\u6298\u3057\u3001\u72ec\u81ea\u306b\u300c\u30bb\u30de\u30f3\u30c6\u30a3\u30c3\u30af\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u300d\u7528\u30e2\u30c7\u30eb\u3092\u69cb\u7bc9\u3059\u308b\u3068\u6c7a\u610f\u3057\u3066\u304b\u3089\u5e7e\u661f\u971c\uff08\u5b9f\u306f4\u30f6\u6708\u7a0b\u5ea6\uff09\u3001\u3084\u3063\u3068\u3053\u3053\u307e\u3067\u6765\u305f\u304b\u3068\u611f\u6168\u6df1\u3044\u3067\u3059\u3002<\/p>\n\n\n\n
\u6b21\u56de\u306f\u3044\u3088\u3044\u3088\u300c\u30d9\u30af\u30c8\u30eb\u5316\u3057\u305f\u4e0a\u3067\u306e\u691c\u7d22\u7cbe\u5ea6\u306b\u95a2\u3057\u3066\u78ba\u8a8d\u3057\u3066\u307f\u308b\u300d\u306b\u30c1\u30e3\u30ec\u30f3\u30b8\u3057\u305f\u3044\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u524d\u56de\u306e\u6295\u7a3f\uff08Mac+PostgreSQL\u3067\u30d9\u30af\u30c8\u30eb\u691c\u7d22\uff09\u3067\u3001\u300c\u6b21\u56de\u306f\u3001\u5148\u306b\u751f\u6210\u3057\u305f\u30e2\u30c7\u30eb\u3092\u4f7f\u7528\u3057\u3066\u30e9\u30d9\u30eb\u90e8\u5206\u3092\u30c8\u30ea\u30df\u30f3\u30b0\u3057\u3001\u30d9\u30af\u30c8\u30eb\u5316\u3057\u305f\u4e0a\u3067\u306e\u691c\u7d22\u7cbe\u5ea6\u306b\u95a2\u3057\u3066\u78ba\u8a8d\u3057\u3066\u307f\u308b\u300d\u3068\u98a8\u5442\u6577\u3092\u5e83\u3052\u307e\u3057\u305f\u304c\u3001\u3044\u3056\u4f5c\u696d\u3092\u59cb\u3081\u3066\u307f\u308b\u3068 […]<\/p>\n","protected":false},"author":6,"featured_media":5697,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":"","_wp_rev_ctl_limit":""},"categories":[210],"tags":[],"class_list":["post-5672","post","type-post","status-publish","format-standard","has-post-thumbnail","category-210"],"_links":{"self":[{"href":"https:\/\/www.wetch.co.jp\/note\/wp-json\/wp\/v2\/posts\/5672","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wetch.co.jp\/note\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wetch.co.jp\/note\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wetch.co.jp\/note\/wp-json\/wp\/v2\/users\/6"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wetch.co.jp\/note\/wp-json\/wp\/v2\/comments?post=5672"}],"version-history":[{"count":0,"href":"https:\/\/www.wetch.co.jp\/note\/wp-json\/wp\/v2\/posts\/5672\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wetch.co.jp\/note\/wp-json\/wp\/v2\/media\/5697"}],"wp:attachment":[{"href":"https:\/\/www.wetch.co.jp\/note\/wp-json\/wp\/v2\/media?parent=5672"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wetch.co.jp\/note\/wp-json\/wp\/v2\/categories?post=5672"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wetch.co.jp\/note\/wp-json\/wp\/v2\/tags?post=5672"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}