No abstract

Small target detection has always been a difficult task in UAV target acquisition. The detection accuracy is low due to the high proportion of small targets, severe mutual occlusion, low feature information and the presence of complex background in the target area. To solve this problem, we propose a high-accuracy small target detection model called YOLOv7X+, which aims to improve the accuracy of UAV small target detection and strike a balance between operational efficiency. The model effectively mitigates the problem of severe mutual occlusion of small targets by introducing the Conv2Formers module for finer feature extraction of spatial information. At the same time, we propose a Bi-level routing attention mechanism based on cavity convolution, which controls feature extraction and delivery at different levels while expanding the perceptual field, enhancing the model's ability to understand and detect the connections between dense small targets through pixel-level processing, making the model more robust in detecting dense small targets in complex and variable scenes, and retaining important background information to increase discriminative accuracy. Finally, the model also is taken into account the multi-dimensional input data under the UAV by adaptively adjusting the size of the convolution kernel to make the algorithm more adaptable to the actual scene requirements. Under the validation test IOU of 0.5, the YOLOv7X+ model achieves an average mean accuracy (mAP50) of 60.3%, which is 5.4% improvement in mAP compared to the original YOLOv7X model and outperforms five current advanced detection methods. The experimental data shows that the model can effectively handle small target detection tasks in complex scenarios and achieve a balance between detection accuracy and efficiency.