A small target detector design for aerial scenarios based on multi-cross adaptive fusion mechanism and high-efficiency feature extraction model

With the rich development of drone technology and artificial intelligence, target detection from the perspective of unmanned aerial vehicles UAV has become a valuable area of research for a large variety of applications. However, detecting small and densely distributed objects with significant overlap in size, all while dealing with inadequate lighting conditions, remains substantial challenges for target detection in this perspective. In this paper, we proposed MPB-YOLO, a high-efficiency small target detection algorithm based on the YOLOv8s model, aiming to overcome the challenges of target detection from the UAV perspective. To address the issue of small target sizes, we have made improvements to the neck structure. Specifically, we introduced a small target detection head on top of the large-scale feature maps and proposed a multi-scale adaptive fusion neck structure. This modification significantly enhanced the model's performance in the presence of small-sized targets. To tackle the challenges of densely distributed and heavily overlapping targets, we incorporated deformable convolutions and designed a spatial coordinate attention mechanism. This mechanism was integrated into the feature extraction module to enhance the network's perceptual capabilities, further improving the model's overall performance. The proposed MPB-YOLO methods had been evaluated in ablation experiments and compared with other state-of-the-art algorithms on VisDrone2019 dataset. The results demonstrate that our MPB-YOLO outperforms other baseline methods in the accuracy of object detection. Compared to YOLOv8s, our method achieved a significant improvement in mAP50 metrics, with a 26.3 increase on the VisDrone2019-test and a 28 increase on the VisDrone2019-val. Additionally, the parameter of MPB-YOLO achieved a 16.8 decrease than the benchmark model. These experiments validated the effectiveness of the MPB-YOLO methods in the task of object detection in aerial scenarios.