국내기사
약지도 학습을 활용한 AI 기반 한국 음식 재료 인식 의미론적 분할 모델 : 불고기 사례 연구 = AI-based weakly supervised semantic segmentation for Korean food ingredient recognition : bulgogi case study
This study investigated the use of weakly supervised learning (WSL) and partial annotation-based semantic segmentation for recognizing ingredients in the traditional Korean dish bulgogi. A dataset was created to encompass various cooking stages and imaging conditions, with pixel-level labels generated for major ingredients: beef, onion, green onion, carrot, chili pepper, mushroom, button mushroom, and king oyster mushroom, using partial annotations. To enhance model robustness, data augmentation techniques such as rotation, scaling, horizontal flipping, and color jittering were employed. The DeepLabV3+ architecture was utilized, with ResNet50 and ResNet101 serving as backbone networks. The results demonstrated that ResNet50 provided stable performance with lower computational costs, while ResNet101 achieved higher segmentation accuracy for smaller or visually complex ingredients. Models trained with data augmentation showed improved recall and F1-scores, especially for smaller ingredient classes. Overall, both backbone models exhibited consistent performance across key segmentation metrics, including mean Intersection over Union (mIoU), precision, recall, and F1-score. These findings indicate that WSL, in conjunction with partial annotation, can effectively facilitate ingredient-level segmentation in mixed dishes like bulgogi.