지붕 개량용 MATT 컬러강판 표면 질감 분류를 위한 딥러닝 기반 자동 검사 기법 연구

초록

This study proposes a slice-based ensemble classification method to effectively detect subtle and spatially distributed texture variations on the surface of MATT color-coated steel. The proposed method divides each input image into multiple spatial slices, applies a shared CNN classifier to each slice, and integrates the predictions using ensemble strategies such as soft voting and top-k voting. This region-based approach enhances robustness by aggregating localized predictions while maintaining model simplicity, making it suitable for industrial applications where data is limited. To optimize performance, the methodology incorporates both Data-Centric and Model-Centric AI strategies. Outliers are removed using Mahalanobis distance computed from image embeddings, and stratified sampling is applied based on class-wise distributions to ensure balanced and diverse training data. OpenCV-based feature enhancement, including Canny edge overlay, is employed to emphasize fine texture boundaries. Additionally, Bayesian hyperparameter optimization is performed to improve model generalization. The proposed method demonstrates high effectiveness in detecting subtle surface texture changes and offers a practical and interpretable solution for automating surface quality inspection in steel manufacturing environments.