Robust Real-Time Pig Detection for Commercial Swine Barns Using an Improved MSCA-RTDETR Model

Pig detection in intelligent livestock farming is challenging due to the difficulty of jointly capturing global context and multi-scale features in complex environments. To address this, we propose MSCA-RTDETR (Multi-scale Content-Aware Real-Time Detection Transformer), which enhances the Real-Time Detection Transformer (RTDETR) architecture with two complementary components. First, we introduce a Content-Aware Token Selection (CATS) backbone that uses content-aware weighting and Top-K sparse attention to efficiently model long-range dependencies and global context. Second, we design a Multi-scale Interaction Residual Block (MIRB) that employs parallel convolutional kernels (3×3 and 5×5) to capture fine-grained local details and broad contour information, handling scale variations due to different viewing distances and pig sizes. On a custom dataset of 8070 images (6955 training, 1115 testing) collected from a commercial pig farm, MSCA-RTDETR outperforms existing detectors, improving AP, AP50, and AP75 by 0.6%, 0.3%, and 0.9% respectively over the strong RTDETRv1 baseline. The model demonstrates effective detection on our self-built dataset, offering a practically viable and accurate solution for intelligent livestock farming.