Development of a Deep Learning-Based Non-contact System for Sow Farrowing Status Classification Using Multiple Instance Learning

Timely detection of farrowing in sows is important for effective farm management and animal welfare. However, existing farrowing monitoring approaches are largely contact-based, which limits their practical applicability in farm environments due to constraints such as cumulative equipment costs and animal stress. To address these limitations, this study developed a deep learning-based non-contact farrowing classification system for sows. The proposed method employed precise region of interest (ROI) cropping based on the Segment Anything Model (SAM) to reduce background interference and consistently include farrowing-related regions. In addition, multiple-instance learning was integrated into a Convolutional Neural Network (CNN)-based classification framework to better aggregate region-wise discriminative cues. Experimental results showed that the proposed method achieved the best overall performance among the compared models, with 85.47% recall and 85.68% F1-score. Compared with the original full-image input, it improved recall by 3.59 percentage points and F1-score by 3.71 percentage points. These results indicate that precise ROI cropping and multiple-instance learning jointly improve farrowing classification performance by reducing background interference and better aggregating region-wise discriminative cues.