SAR image ship detection based on light-weight multi-scale feature enhancement network

Abstract

Aiming at the problem of insufficient detection performance due to complex background and small-scale characteristics of the target in SAR-based image ship detection applications, this paper constructs a lightweight multi-scale feature enhancement network based on deep learning algorithms for ship detection. The network first extracts the ship features in the image by using the light-weighted and improved VGG-16 network, and designs a two-way feature fusion module to forward fuse and reverse fuse the low-level detail features and high-level semantic features respectively to achieve multi-scale feature aggregation; then introduces a convolutional block attention module to enhance the network's ability to extract ship features in complex scenes from the perspective of both channel features and spatial features. Further, the K-means clustering algorithm is used to optimize the anchor scale design based on the real border information of ship targets to enhance the classification and localization accuracy of the network for multi-scale ship targets; finally, the Soft-NMS algorithm is introduced to improve the leakage detection problem in the densely distributed ship scenes. The detection performance of the network is verified using the SAR image ship dataset SSDD, and the research results show that the detection accuracy of the network is 95.12% and the detection speed is 18.7 FPS, which is 6.88% higher than that of the Faster R-CNN network in terms of detection accuracy and more than 2 times higher in terms of detection efficiency, effectively improving the ship detection performance of small-scale targets in complex backgrounds. It can meet the demand of high accuracy and real-time detection of ship targets in SAR images.