Achieving Ship Target Detection in Radar Images Via Dual-Route Feature Extraction and Adjacent-Layer Feature Fusion

Focusing on the problem of balancing computational cost and detection accuracy in marine SAR image target detection, we propose a lightweight method based on Dual Route Feature Extraction and Adjacent-Layer Feature Fusion, which mainly consists of two steps. Firstly, we designed a dual-route feature extraction backbone, input features are equally divided in channel dimension, and feature extraction is carried out separately, which reduces the computational parameters. Meanwhile, to address the diversity of ship scales and rotation angles on the marine surface, deformable convolution is incorporated into the backbone, skip connection is used to enhance the information of small ship targets. Subsequently, in the feature fusion part, in order to reduce the damage caused by non-linear operations to the hierarchical correlation of the feature pyramid, a strategy of Adjacent-Layer Feature Fusion is adopted, and the feature output with a pyramidlike structure is generated. Experiments have proven that the constructed structure can maintain detection accuracy and reduce the computational cost of the network. The publicly available SSDD and HRSID are used to verify the advantage of the network. The experiments show that the network structure designed can effectively reduce the computational cost of the network while maintaining a relatively high detection accuracy.