SALC-Net: A lightweight contour-preserving segmentation network for yak body segmentation in complex grazing environments.

Abstract

Accurate livestock segmentation is a key prerequisite for non-contact image-based analysis and intelligent pasture management, yet remains challenging on resource-constrained edge devices. Although lightweight networks are suitable for real-time deployment, they often suffer from limited geometric adaptability and insufficient boundary preservation, which reduces the reliability of downstream shape-related analysis for non-rigid livestock targets. To address this issue, we propose SALC-Net, a lightweight segmentation framework for yak body contour extraction. SALC-Net combines a re-parameterized MobileNetV2 backbone for efficient inference, a Scale-Adaptive Efficient Dynamic Pyramid (SA-EDP) module for low-cost adaptive receptive-field modeling, and a Linear Cross-Scale Fusion (LCSF) module for contour-preserving feature reconstruction. Experiments on a custom high-altitude yak dataset show that SALC-Net achieves 93.37% mIoU at 129 FPS, demonstrating a favorable trade-off between segmentation accuracy and real-time efficiency.