SAR Ship Detection with Large-Kernel Convolution and Channel Weighted Multi-Scale Feature Fusion

Ship detection using Synthetic Aperture Radar (SAR) images is indispensable for maritime transportation. However, this task still faces several challenges such as low contrast image, limited pixel occupation, dense ship arrangement and irregular background clutter. Existing deep learning based models which are designed for general object recognition have limited generalization capability when trained on SAR datasets. To address the above issues, this paper develops an improved YOLOv8-based model. In the feature extraction stage, light-weight stem and large kernel convolutional layers are employed to increase effective receptive field. Re-parameterized technique is applied to maintain a low inference cost of the large kernel. In the feature fusion stage, channel-weighted feature concatenation is proposed to differentiate importance between multi-scale feature maps. In the detection head, an extra detection head from early-stage convolutional layers is added. Shared convolution and group normalization are applied to standardize the model output, ensuring consistent performance. The proposed model is evaluated using three public datasets. It achieves Average Precision of 94.9% on the HRSID, 98.7% on the SSDD, and 77.2% on the LS-SSDD-v1.0. With a parameter size of 10M, the proposed model is suitable for real-time SAR ship detection in challenging surveillance scenarios.