AUTHOR=Xie Hongtu , Jiang Xinqiao , Zhang Jian , Chen Jiaxing , Wang Guoqian , Xie Kai 

TITLE=Lightweight and anchor-free frame detection strategy based on improved CenterNet for multiscale ships in SAR images

JOURNAL=Frontiers in Computer Science

VOLUME=Volume 4 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/computer-science/articles/10.3389/fcomp.2022.1012755

DOI=10.3389/fcomp.2022.1012755

ISSN=2624-9898

ABSTRACT=Ship detection using the synthetic aperture radar (SAR) images has the important applications in the military and civilian fields, but the ship different sizes would degrade the detection accuracy of the multiscale ship. Aiming at the problem of the poor accuracy and low efficiency of the multiscale ship detection in the complex scenes, this paper has proposed a lightweight and anchor-free frame detection strategy of the multiscale ship in the SAR images. First, to deal with the problems of the limited training samples, different sizes, attitudes and angles of the ships in SAR images, a data augmentation strategy suitable for the SAR images is adopted to expand the training space, and then the multiscale training is introduced to enhance the model generalization ability for the multiscale ship detection. Second, a lightweight and anchor-free ship detection model based on the improved CenterNet is proposed, which abandons the dense anchor frame generation and extracts the key point of the ships for the detection and positioning. Compared with the anchor frame-based detection method, this proposed detection model does not need to use the post-processing method to remove the redundant anchor frames, which can accurately locate the center point of the ships and has the better detection performance. Third, to reduce the model size and simplify the model parameters, the more lightweight network design is adopted in combination with the characteristics of the SAR images. Hence, a residual network (ResNet) with the fewer convolutional layers is constructed as the backbone network, and the cross-stage partial network (CSPNet) and spatial pyramid pooling (SPP) network are designed as the bottleneck network. The shallow ResNet can fully extract the SAR image features and reduce the training overfitting, and CSPNet and SPP can effectively combine the low-level image features to obtain the high-level features, reducing the model computation and enhancing the feature extraction ability. Finally, the evaluation index of the common objects in context dataset is introduced, which can provide the higher-quality evaluation results for the ship detection accuracy and provide the comprehensive evaluation indicators for the multiscale ship detection.