Accurate livestock detection under dense aggregation and occlusion is essential for smart agriculture tasks such as farm management, disease prevention, and automated monitoring. Real-world constraints—high object density, severe inter-individual occlusions, complex backgrounds, and variable postures—degrade feature integrity and reduce detection reliability. Traditional handcrafted methods and some earlier deep-learning solutions struggle with occluded, overlapping instances or incur high computational overhead that limits edge deployment. To address these problems, the authors propose DOLD-Net, a framework that applies progressive global-to-local refinement to preserve discriminative details while maintaining coherent global structure.
DOLD-Net is organized into three tightly coupled stages: Backbone, Neck, and Head. The Backbone focuses on robust feature extraction under aggregation. The Neck performs semantic propagation and anti-occlusion refinement across scales. The Head converts restored multi-scale features into final detections using the reported D-FINE detection architecture. The design intent is to resolve occlusion-induced semantic discontinuities while balancing accuracy and computational efficiency.
The Backbone implements a Dual-Branch Occlusion-Aware Network (DBOAN) to overcome limitations of single-view feature encodings in crowded scenes. DBOAN concurrently models local gradient flows and global topological structures so that fine-grained details and spatial relationships are jointly preserved. This collaborative learning mechanism aims to reduce feature confusion caused by overlapping instances and provide a stronger initial representation for subsequent refinement stages.
In the Neck stage, the authors introduce Context-Guided Focus Propagation (CGFP) to bridge semantic gaps created by occlusion. CGFP constructs a global core representation by extracting key semantic cues via a Cross-Scale Feature Enhancer (CSFE). This global core acts as a top-down prior that propagates discriminative semantic information across feature hierarchies, guiding the restoration and inference of occluded or partially visible targets. The approach differs from simple feature aggregation by explicitly injecting global context to inform local feature recovery.
To address boundary blurring and entangled representations common in dense clusters, DOLD-Net integrates a Frequency–Spatial Anti-occlusion Module (FSAM) within the propagation pathway. FSAM performs frequency-domain decoupling of features to sharpen ambiguous boundaries that may be poorly separated by spatial-only modeling. This frequency–spatial decomposition helps distinguish overlapping objects and mitigates semantic mixing introduced by occlusion.
To ensure consistency across scales and repair local information loss, the network uses a Context-Aware Occlusion Fusion Module (CAOFM). CAOFM performs adaptive, pairwise cross-scale interactions between adjacent feature levels to recover missing details and to correct spatial discontinuities. This module supports the final conversion of refined features into robust detections.
The paper reports using the D-FINE architecture as the detection head to transform the multi-scale, occlusion-corrected features into bounding-box and class predictions. The source describes the overall pipeline as a progressive global-to-local refinement before passing features to the detecti(…) head. (The source text was truncated at that point of the Head description.)
Experiments were carried out on four publicly available livestock datasets: GooseDetect, SheepCounter, ChickenFlow, and CherryChèvre. The source lists DOIs and dataset access locations. Performance is reported using Average Precision (AP) to quantify detection accuracy. The authors state that the implementation and source code are publicly available at https://github.com/Jiakaida/DOLD-Net.
The paper reports that DOLD-Net achieves a favorable balance between accuracy and efficiency. On the four evaluated datasets, reported AP scores are: 57.7%, 67.5%, 52.5%, and 62.5% respectively. These values are presented in the abstract and represent the primary quantitative outcomes highlighted by the authors.
The authors emphasize that many prior approaches improve occluded-object performance by stacking complex modules, which increases parameters and computational cost and limits real-world deployment. DOLD-Net is presented as a design that seeks to balance model complexity with detection accuracy by employing targeted modules—DBOAN, CGFP, FSAM, and CAOFM—that collaborate to restore occluded semantics while aiming to remain efficient for practical use. Specific parameter counts, inference speed, or edge-device benchmarks were not detailed in the provided source excerpt.
DOLD-Net is proposed to improve densely occluded livestock detection via global–local feature collaboration. The architecture integrates occlusion-aware backbone modeling, context-guided semantic propagation, frequency–spatial anti-occlusion processing, and context-aware fusion to recover missing details and sharpen boundaries. Reported AP performance on four public datasets indicates the method's effectiveness. The authors provide code at https://github.com/Jiakaida/DOLD-Net and cite dataset sources (GooseDetect, SheepCounter, ChickenFlow, CherryChèvre). Funding sources and a declaration of no competing interests are reported in the article metadata.