关键词: 航拍图像, 可变形卷积, 小目标检测, 多尺度特征融合, 目标检测层

Abstract:

In order to solve the problem that UAV aerial images have a large number of small target samples but little extractable feature information, which is not conducive to improving the accuracy of aerial target detection, an improved small target detection algorithm for aerial photography based on YOLOv8s is proposed. The algorithm applies deformable convolution to the feature extraction module of the backbone network to adaptively capture the details of the target at different locations and scales. The feature information at different scales of the backbone network is extracted and enhanced by the feature collection module in the multilevel information fusion function block to obtain fine global features.The context-rich semantic information is injected into the small target detection layer using the information fusion module to realize the fusion of local and global information, and the fused features are input into the detection network to obtain the detection results. The experimental results demonstrate that the proposed algorithm improves the recognition average accuracy and recall by 14.96% and 10.85%, respectively, compared to the baseline model. The leakage rate is reduced by 7.11%. The improved target detection algorithm has the highest average accuracy for small target detection compared to mainstream detection algorithms.

Key words: aerial image, deformable convolution, small target detection, multi-scale feature fusion, object detection layer