Multiscale feature cross-layer fusion remote sensing target detection method

Target detection based on remotely sensed images, which has recently attracted much attention, is a fundamental but challenging task. In remote sensing images, the problem of difficult recognition of small targets or targets with a large aspect ratio arises because the targets have the characteristics of small proportion, dense distribution, and multidirectionality. To address the above problems, this article proposes an improved multiscale feature cross-layer fusion remote sensing target detector based on YOLOv5. First, this method introduces the circular smooth label technique, using YOLOv5 as a rotation detector to solve the angular boundary condition and angle prediction problem for large aspect ratio targets. Second, the explicit visual centre module is introduced to solve the problem of missed detection in target-dense distribution tasks. Finally, a multiscale feature cross-layer fusion structure (S-160) is proposed based on YOLOv5, which improves the detection accuracy of each scale target by fusing shallow and deep feature information and introduces new large-scale features for small target detection to solve the problem that ultrasmall targets in remote sensing images cannot be recognised. Our experiments were conducted on three public remote sensing datasets, DOTA, DIOR-R, and HRSC2016, and the average accuracy (mAP) on the datasets was 76.50%, 70.34%, and 97.68%, respectively, demonstrating the substantial detection performance of the proposed method.