SALC-Net: Lightweight AI for yak segmentation in complex grazing environments

Researchers have developed SALC-Net, a lightweight AI system that segments yak bodies in complex grazing environments with 93.37% accuracy at 129 frames per second, enabling real-time livestock monitoring on resource-limited edge devices. The system combines a re-parameterized MobileNetV2 backbone with two novel modules: a Scale-Adaptive Efficient Dynamic Pyramid for adaptive receptive-field modeling and a Linear Cross-Scale Fusion module for contour preservation. Tested on 1,550 high-resolution images from Qinghai Province at 2,800-4,500 meters elevation, SALC-Net outperformed state-of-the-art models while reducing parameters by approximately 95.8% compared to standard baselines. The dataset covers challenging conditions including complex backgrounds, occlusion, pose variation, and illumination changes. The authors note limitations including single-view RGB imaging without depth information and plan future work on depth-aware sensing and model quantization for long-term deployment in alpine pastoral environments.