DOLD-Net: Dense occluded livestock detection via global-local feature collaboration

Precise livestock detection is fundamental to smart agriculture; however, real-world environments often present challenges such as high object density, severe occlusion, and ambiguous boundaries. To address feature integrity degradation under such conditions, this paper proposes DOLD-Net: a model for densely occluded livestock detection utilizing global–local feature collaboration. First, this paper develops a Dual-Branch Occlusion-Aware Network (DBOAN) backbone to mitigate feature confusion stemming from object aggregation. The DBOAN models both local gradient flows and global topological structures. By integrating these two aspects, the network preserves fine-grained details while maintaining coherent spatial relationships. To further resolve semantic discontinuities in occluded regions, a Context-Guided Focus Propagation (CGFP) mechanism is designed. The CGFP constructs a global core representation through long-range dependency modeling. This representation propagates discriminative semantic priors in a top-down manner across feature hierarchies. Furthermore, a Frequency–Spatial Anti-occlusion Module (FSAM) decouples entangled representations in the frequency domain to sharpen ambiguous boundaries. Additionally, a Context-Aware Occlusion Fusion Module (CAOFM) compensates for local information loss through adaptive cross-scale feature interactions. Extensive experiments on four public datasets demonstrate that DOLD-Net achieves a superior balance between accuracy and efficiency, attaining AP scores of 57.7%, 67.5%, 52.5%, and 62.5%, respectively. The source code is publicly available at https://github.com/Jiakaida/DOLD-Net.