Focus-Based Multi-Scale Approaches for Tiny Target Identification in Aerial Images

Abstract

Small targets in remote sensing images have too few discriminative features, making them easily confused with background information and difficult to find, which reduces the detection accuracy when detecting aerial images using general target detection networks. To solve the above problems, they proposed a remote sensing small object detection network based on an attention mechanism and multi-scale feature fusion and named it. First, the detection head enhancement module is designed to enhance the representation of small object features by combining multi-scale feature fusion and an attention mechanism. Secondly, the Attention Mechanism Channel Cascade (AMCC) is designed to reduce redundant information in the feature layer and protect small objects from losing information during feature fusion. Then, the regularized Wasserstein distance is introduced and combined with the generalized cross-sum as the position regression loss function to improve the model optimization weights and regression box accuracy for small objects. Finally, a target detection layer is added to improve the ability to extract target features at different scales. Experimental results on the drone dataset VisDrone2021 and a self-made dataset show that AMCC is 2.4% and 3.2% higher than YOLOv5s, respectively, effectively improving the recognition accuracy of small objects.