Cattle segmentation in precision livestock farming faces persistent challenges from background-induced semantic uncertainty and occlusion-driven geometric distortion. To address these issues, this paper proposes MFDSU-Net, a novel U-Net architecture that jointly refines semantic and geometric representations. The MFDSU-Net architecture comprises two core modules, including the Multi-Scale Feature Aggregation Block (MFABlock) and the Multi-scale Deformation Sampling Block (MDSBlock) respectively. Specifically, MFABlock captures contextual information across varying receptive fields, enhancing model robustness against complex backgrounds and scale variations. Concurrently, MDSBlock adaptively models spatial deformations, preserving boundary fidelity against occlusions and pose variations. Experimental results show that the proposed MFDSU-Net achieves a Dice of 88.14% and an IoU of 81.38%, outperforming existing state-of-the-art models by approximately 0.5% and 0.8%. Furthermore, with only 0.7M parameters and 2.8 GFLOPs, MFDSU-Net maintains high inference efficiency, rendering it suitable for real-time deployment on resource-limited agricultural edge devices.
Hao et al. (Wed,) studied this question.