비전 데이터 수집과 전처리

import cv2 import numpy as np from typing import Tuple class ImagePreprocessor: """제조 비전 이미지 전처리""" def __init__(self, target_size: Tuple[int, int] = (224, 224)): self.target_size = target_size def preprocess(self, image: np.ndarray) -> np.ndarray: """전처리 파이프라인 실행""" # 1. 노이즈 제거 image = cv2.GaussianBlur(image, (3, 3), 0) # 2. 히스토그램 평활화 (대비 개선) if len(image.shape) == 3: lab = cv2.cvtColor(image, cv2.COLOR_BGR2LAB) lab[:,:,0] = cv2.createCLAHE(clipLimit=2.0).apply(lab[:,:,0]) image = cv2.cvtColor(lab, cv2.COLOR_LAB2BGR) # 3. 정규화 image = image.astype(np.float32) / 255.0 # 4. 리사이즈 image = cv2.resize(image, self.target_size) return image def align_roi(self, image: np.ndarray, template: np.ndarray) -> np.ndarray: """템플릿 매칭으로 ROI 정렬""" result = cv2.matchTemplate(image, template, cv2.TM_CCOEFF_NORMED) _, _, _, max_loc = cv2.minMaxLoc(result) h, w = template.shape[:2] roi = image[max_loc[1]:max_loc[1]+h, max_loc[0]:max_loc[0]+w] return roi

import albumentations as A from albumentations.pytorch import ToTensorV2 # 학습용 증강 (강한 증강) train_transform = A.Compose([ A.Resize(224, 224), A.HorizontalFlip(p=0.5), A.VerticalFlip(p=0.5), A.RandomRotate90(p=0.5), A.ShiftScaleRotate(shift_limit=0.1, scale_limit=0.2, rotate_limit=45, p=0.5), # 조명 변동 시뮬레이션 A.RandomBrightnessContrast(brightness_limit=0.3, contrast_limit=0.3, p=0.5), A.RandomGamma(gamma_limit=(80, 120), p=0.3), A.GaussNoise(var_limit=(10, 50), p=0.3), # 블러 (초점 불량 시뮬레이션) A.OneOf([ A.MotionBlur(blur_limit=5), A.GaussianBlur(blur_limit=5), ], p=0.3), A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]), ToTensorV2(), ]) # 검증/추론용 (증강 없음) val_transform = A.Compose([ A.Resize(224, 224), A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]), ToTensorV2(), ])

import os import hashlib from PIL import Image from collections import Counter class DatasetValidator: """데이터셋 품질 검증""" def validate(self, data_dir: str) -> dict: """데이터셋 검증 실행""" issues = { 'corrupted': [], 'duplicates': [], 'size_outliers': [], 'class_imbalance': None } hashes = {} sizes = [] labels = [] for root, dirs, files in os.walk(data_dir): for file in files: if not file.lower().endswith(('.jpg', '.png', '.bmp')): continue path = os.path.join(root, file) label = os.path.basename(root) labels.append(label) # 1. 손상된 이미지 검사 try: img = Image.open(path) img.verify() img = Image.open(path) sizes.append(img.size) except: issues['corrupted'].append(path) continue # 2. 중복 이미지 검사 with open(path, 'rb') as f: file_hash = hashlib.md5(f.read()).hexdigest() if file_hash in hashes: issues['duplicates'].append((path, hashes[file_hash])) else: hashes[file_hash] = path # 3. 크기 이상치 검사 avg_size = np.mean([s[0]*s[1] for s in sizes]) for i, size in enumerate(sizes): if size[0]*size[1] < avg_size * 0.5: issues['size_outliers'].append(i) # 4. 클래스 불균형 검사 class_counts = Counter(labels) issues['class_imbalance'] = dict(class_counts) return issues # 사용 예시 validator = DatasetValidator() issues = validator.validate('defect_dataset/processed/train') print(f"손상 이미지: {len(issues['corrupted'])}개") print(f"중복 이미지: {len(issues['duplicates'])}개") print(f"클래스 분포: {issues['class_imbalance']}")