Military Aircraft Recognition Method Based on Attention Mechanism in Remote Sensing Images

Abstract

Remote sensing images play a crucial role in fields such as reconnaissance and early warning, intelligence analysis, etc. Due to factors such as climate, season, lighting, occlusion and even atmospheric scattering during remote sensing image acquisition, targets of the same model exhibit significant intra-class variability. This article applies deep learning technology to the field of military aircraft recognition in remote sensing images and proposes a You Only Look Once Version 8 Small (YOLOv8s) remote sensing image military aircraft recognition algorithm based on an attention mechanism—YOLOv8s-TDP (YOLOv8s+TripletAttention+dysample+PIoU). First, the TripletAttention attention module is used in the neck network, which captures cross-dimensional interactions and utilises a three-branch structure to calculate attention weights. This further enhances the network's ability to preserve details and restore colours in the process of image fusion. Secondly, an efficient dynamic upsampler, dysample, is used to achieve dynamic upsampling through point sampling, which improves the problems of detail loss, jagged edges, and image distortion that may occur with nearest neighbour interpolation. Finally, replacing the original model loss function with PIoU (Pixels Intersection over Union), IoU (Intersection over Union) is calculated at the pixel level to more accurately capture small overlapping areas, reduce missed detection rates, and improve accuracy. On the publicly available dataset The Remote Sensing Image Military Aircraft Target Recognition Dataset(MAR20), our proposed YOLOv8s-TDP model achieved a $Precision$ of 82.96%, $Recall$ of 80.71%, $mA{P}_{0.5}$ of 87.11% and $mA{P}_{0.5-0.95}$ of 65.88%, outperforming the original YOLOv8s model, RT-DETR model, YOLOv5 series model, YOLOv7 series model, and YOLOv11 series model. Compared with the original YOLOv8s model, the YOLOv8s-TDP model improves $Precision$ by 0.23%, $Recall$ by 2.61%, $mA{P}_{0.5}$ by 2.76%, and $mA{P}_{0.5-0.95}$ by 2.49%, verifying that the algorithm has good performance in remote sensing image military aircraft recognition.