关键词: 小目标检测, 特征融合, 空洞卷积, 无人机图像, Swin Transformer

Abstract: Aiming at the problems of complex target scenes, diverse scales, dense small targets and severe occlusion in UAV aerial images, a multi-scale dilation convolution Swin-Det UAV image target detection algorithm was proposed. Firstly, Swin Transformer is used as the backbone feature extraction network, and a spatial information blending module is introduced into the backbone network to solve the problem of blurred target information due to occlusion between objects; secondly, a fusion of dilation feature pyramid network is proposed, which uses multiple branches to the dilation convolution fuses the feature information, effectively improves the receptive field of the network and the reuse of feature information, so that the model can learn detailed features of different dimensions; finally, linear interpolation method and multi-task loss function are used to solve the problem of inconsistency in the prediction area. Matching and sample imbalance problems improve the detection accuracy of the model. Experimental results on the VisDrone data set show that the mean average precision of the proposed algorithm reaches 27.2%. Compared with the original Swin Transformer, the mean average precision is improved by 4.1 percentage points and converges faster under the same training batch. The proposed algorithm can achieve high-precision detection of drone image targets in complex scenarios.

Key words: small-target detection, feature fusion, dilation convolution, UAV images, Swin Transformer