MBAF-MSCNet: Multi-Branch Fusion for Multi-Scale Context in Infrared Wildlife Detection

This paper proposes MBAF-MSCNet, an attention-enhanced multi-branch fusion network for infrared wildlife detection in Unmanned Aerial Vehicle (UAV) imagery, addressing critical challenges, including small target ambiguity, complex background interference, and target occlusion. Based on YOLOv11-n, the model integrates three synergistic innovations: a Parallelized Block-Aware Attention Enhancement module (PPA_Enhancer) that extracts multi-scale features through local-global branches while optimizing channel dimensionality; a Dilated Residual Module (DWR) employing depth-separated dilated convolutions and a two-step residual mechanism (regional-semantic residualization) for efficient multi-scale context capture; and angle-sensitive SIoU loss incorporating geometric constraints (angle/distance/shape) to refine bounding box regression. Based on the BIRDSAI dataset (18,557 training images), MBAF-MSCNet achieved 93.6% mAP50 and 50.2% mAP50-95, surpassing YOLOv11-n by 2.1% in mAP50while maintaining real-time inference (256.4 FPS) with minimal parameters (3.279M). Cross-dataset validation of ISOD further demonstrated exceptional generalizability (79.6% AP50). The superiority of the model lies in its balanced architecture, which significantly enhances the detection accuracy under complex infrared conditions without compromising computational efficiency.