Due to the scattering effect of synthetic aperture radar (SAR) and the resolution limitation of wavelength and antenna size, it is difficult to obtain the details and boundary information of small-sized targets, which affects the detection accuracy. To improve the accuracy of SAR ship detection and reduce the false detection rate, SAR ship detection based on an efficient multiscale feature enhancement network (EMFENet) is proposed. First, in the backbone network, a multifeature fusion structure is used to construct a step-by-step residual partial convolution block as the main feature extraction module, which improves the recognition ability of ship targets and effectively reduces the number of parameters. Second, the feature fusion part constructs an efficient layer convolution block with a spatial channel attention mechanism and an aggregation structure and pays more attention to the geometric information of the model. The multiscale fusion feature pyramid network is constructed using the receptive field block, and the multiscale features are further fused to enhance the feature capture ability of the overall structure of the network. Finally, the normalized Gaussian Wasserstein distance metric is introduced into the detection head loss regression function to enhance the detection ability of small targets. The effectiveness of the proposed network was compared with mainstream detection algorithms using the high-resolution SAR images dataset (HRSID) and the SAR ship detection dataset (SSDD). Experiments showed that the proposed network has significant advantages in improving the detection accuracy of SAR ships and in reducing false and missed detections.
Synthetic Aperture Radar Ship Detection Based on Efficient Multiscale Feature Enhancement Network
IEEE Transactions on Aerospace and Electronic Systems ; 60 , 6 ; 8212-8225
2024-12-01
25614177 byte
Article (Journal)
Electronic Resource
English
Extreme Learning Machine Based Ship Detection Using Synthetic Aperture Radar
Springer Verlag | 2018
|Ship Detection Using RADARSAT-1 Synthetic Aperture Radar: Validation with VTS
British Library Conference Proceedings | 2007
|