RAR-LbC 기반 모돈 분만 여부 분류 시스템 개발

초록

Timely identification of sow farrowing is important for farm management and piglet welfare. This study proposes RAR-LbC, a loss-based ROI-aligned training strategy for classifying top-view farrowing-pen images into farrowing and non-farrowing classes. EfficientNetV2 is used as the backbone, and the proposed method improves training by modifying the loss function without changing the network architecture. RAR-LbC combines ROI-based Activation Regularization (RAR), which suppresses activation probability mass outside a predefined ROI, with Localization by Classification (LbC), which aligns the attention distribution used for feature aggregation with the ROI. Experiments showed that RAR-LbC improved overall classification performance compared with the original-image baseline, increasing recall from 72.58% to 74.42%, which indicates improved detection of farrowing frames. Paired t-test, ablation, and sensitivity analyses further confirmed the effectiveness of the proposed ROI-aligned loss design. These results suggest that RAR-LbC can improve sow farrowing status classification by guiding the model toward farrowing-relevant regions during training.

키워드