A method for herder sheep ownership identification based on an improved Mask2Former

Abstract

Overgrazing is one of the primary causes of ecological degradation on the Qinghai Plateau, which severely restricts both the sustainable use of grasslands and the effective management of sheep herds. To address this challenge and achieve a balance between livestock and grassland resources, a semantic segmentation model based on the color features of sheep backs is proposed to accurately identify individual sheep and determine their herder affiliation. To support model training, a dedicated dataset was constructed for sheep back color segmentation and herder classification in the Qinghai Plateau region, offering a rich and diverse set of samples. In terms of model improvement, the proposed method builds upon the original Mask2Former network by introducing a Feature Pyramid Network (FPN), Haar wavelet transform, and a Convolutional Block Attention Module (CBAM). These components enhance the model's performance in complex backgrounds and fine-grained segmentation tasks by optimizing multi-scale feature fusion, improving local feature extraction, and focusing on key regions. Experimental results show that, compared with the original Mask2Former, the improved model achieves increases of 1.89%, 1.26%, 1.19%, and 1.22% in mIoU, Precision, Recall, and F1-score, respectively. These enhancements significantly improve the model's accuracy in fine-grained color segmentation. They also demonstrate its robustness and broad applicability in complex environments. This study provides an innovative solution for sheep management and herder attribution on the Qinghai-Tibet Plateau and opens a new research direction for color-feature-based image segmentation tasks.