Entropy-enhanced swin transformer for small object detection in UAV images

Small object detection in Unmanned Aerial Vehicle (UAV) images is challenging due to the few object pixels including only limited semantic information and disturbed by the background. Moreover, the spatial features of small objects tend to gradually disappear as the depth of deep neural networks increases. To address the issue, this paper proposes an Entropy-Enhanced Swin Transformer (EEST) algorithm. In this method, the Value matrix of Swin Transformer is constituted using image entropy that is the prior knowledge produced by gray-level co-occurrence matrix (GLCM). The EEST is divided into 4 different GLCM Entropy Swin Transformer blocks. The shallow block can extract the discriminative features guided by the complexity of pixel intensity distribution, and the deep block can enhance the faded spatial features by summarizing the variation of the extracted feature map to improve the small object detection performance. The effectiveness of the proposed algorithm is demonstrated by comparative experiments on UAV-DA and VisDrone datasets. On the UAV-DA dataset, the proposed method reaches the highest (Formula presented.) (68.9%) and (Formula presented.) (29.3%); on the VisDrone dataset, EEST achieves the best performance across all metrics: mAP (27.7%), (Formula presented.) (45.9%), (Formula presented.) (18.0%), and (Formula presented.) (40.0%).