MSS-Net: a lightweight network incorporating shifted large kernel and multi-path attention for ship detection in remote sensing images

Abstract

For the challenges with nearshore small ship detection in remote sensing images (RSIs) under complex background, a lightweight network called “multi-path attention and shifted large kernel network for ship detection” (briefly called “MSS-Net”) in RSIs is proposed. Firstly, shifted large kernel with feature enhancement module (SLKE) is developed to enlarge receptive field by decomposing large kernels and shift operation while performing dynamic channel attention. Secondly, multi-path attention (MPA) is designed to effectively retain the co-calibration of spatial-channel information of ships. Thirdly, shared convolutional detection head (SCDH) is built to unify multi-scale features, reducing parameter redundancy. The proposed MSS-Net is validated through three public datasets, TGRS-HRRSD, MASATI and LEVIR. Using YOLOv8 as a baseline model for comparison analysis. The results demonstrate that the mAP50 reaches 97.5%, 78.8%, and 93.2% with the three datasets, respectively. The mAP50 with the proposed MSS-Net is higher 2.9% than YOLOX, 4.4% than RetinaNet in popular one-stage ship detection models, and 4.8% than Faster R-CNN; 2.7% than Cascade R-CNN in two-stage ship detection models. Moreover, the parameters in the MSS-Net reduces 26.7% relative to the baseline model, achieving a lightweight design. Besides, ablation experiments are conducted with the TGRS-HRRSD dataset. The results demonstrates that the SLKE increases mAP50 by 1.1%, the MPA increases mAP50 by 1.7%, while the SCDH reduces parameters by 35%. These results demonstrate that the MSS-Net achieves notable advances for lightweight ship detection.