DIA-YOLO: Diffusion-Based Involution Attention YOLO for Overhead Imagery

In this paper, a Diffusion-based Involution AttentionYOLO network and a Diffusion-based Four-channel Fusion module are proposed, both specifically for oriented bounding box in overhead imagery. The propoased fusion module uses the Marigold depth-estimation network to infer depth information from a single RGB image, producing a four-channel RGBD input. In the proposed network backbone, involution captures fine-grained spatial relationships across color and depth patterns, and the Convolutional Block Attention Module injects the most salient features and locations into the detection head. Experimental results demonstrate significant improvements across multiple datasets. On PCB data, DIA-YOLO achieves 91.6% improvement in mAP50-95 (32.0% to 61.3%) with 16.5% faster inference time (0.71s). Cross-dataset evaluation shows consistent improvements: 21.7% in <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">PCB_self</i>, 11.8% in <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">PCB_open</i>, 31.3% in DOTA, and 7.6% in AITOD. Comparative analysis confirms superiority with 79.32% mAP50 onDOTAand 89.9% mAP50 on AITOD. As the inference time is 0.71s—well below 1 second—this method satisfies real-time requirements for practical overhead imagery applications, providing an effective solution for PCB inspection, satellite analysis, and autonomous systems.