改进YOLOv10的遥感图像轻量级小目标检测算法OA

Remote sensing images have limited object sizes and varying object scales,so it is difficult to detect small objects.This leads to low detection accuracy and slow detection speed.In light of this,the paper proposes a lightweight small object detection algorithm SHAM-YOLO-EfficientViT based on improved YOLOv10.EfficientViT was used as the backbone,which enriches image features while reducing computational load.In the Neck layer,the spatial-channel decoupling downsampling(SCDown)was replaced with spatial-depth decoupling downsampling(SPDown)to compress model size while preserving crucial downsampling information.Furthermore,a scale sequence feature fusion and hybrid attention mechanism(SHAM)was designed to adaptively focus on the key features of small objects in remote sensing images and maintain multi-scale feature information.Additionally,a quadruple downsampled branch was incorporated to enrich feature representations and enhance the performance of object detection.The experiments show that SHAM-YOLO-EfficientViT achieves an mAP@0.5 of 42.8%on the VisDrone2019 dataset,which is 4.4%higher than that of the baseline model YOLOv10;on the ROSD dataset it reaches 98.8%,a gain of 1.0%compared to that of the baseline.Additionally,the number of model parameters is reduced by 11×106,which indicates that the proposed method possesses better generalization capabilities.