Recognition Method of Ship Targets for SAR Based on M3Net

To address the challenges of significant intra-class variations and high inter-class similarities in ship target recognition within synthetic aperture radar (SAR) images, this paper proposes a novel recognition method based on a multibranch, multi-information, multi-depth feature fusion complex-valued network (M3Net). Traditional methods predominantly rely on manually designed amplitude features, failing to fully exploit the inherent complex-valued nature of raw SAR data and neglecting the crucial phase information and its coupling relationship with amplitude. This limitation results in insufficient characterization of ships’ fine structures and ultimately restricts recognition accuracy and model generalization capability. Through in-depth analysis of the noncircularity and complex signal kurtosis characteristics of ship targets, this study reveals that these features can effectively characterize the scattering properties distinguishing ships from the sea background, highlighting the representational advantages of complex-domain statistics for ship scattering characteristics. Building on this foundation, a deep complex feature extraction module (CFEM) is designed. This module employs complex-valued convolutional operations to extract amplitude-phase coupled features and innovatively introduces a cross-fusion of real and imaginary activation (CRIA) function. The CRIA mechanism, utilizing a dual-activation function cross-coupling approach, achieves nonlinear feature interactions and enhances the representational capacity for complex-valued features. Furthermore, the multi-branch, multi-information, multi-depth fusion network M3Net is constructed. M3Net synergistically in- tegrates a core complex-valued convolutional neural network (CV-CNN) backbone, a pre-trained CFEM branch, and a real-valued feature branch. By incorporating a complex-domain attention mechanism, M3Net achieves dynamic weighted fusion of these heterogeneous features, adaptively highlighting the most discriminative feature channels. Experimental results on the reconstructed OpenSARship dataset demonstrate the effectiveness of the proposed method. Compared to the traditional CV-CNN, our approach achieves a 5.89% improvement in overall accuracy and reduces the maximum accuracy deviation across classes to 6.82%, significantly enhancing category balance.