A multivariate calibration framework with global-local interaction and edge-aware enhancement for sonar image despeckling

During sonar image acquisition in underwater environments, speckle noise severely degrades image quality, hindering reliable interpretation for critical tasks such as infrastructure monitoring and post-disaster assessment. To address this challenge, we propose a novel multivariate calibration architecture to overcome the limitations of existing methods for sonar image despeckling, including the reliance on single-network architectures and the neglect of inter-image correlations. By refining the output based on imaging principles and introducing a multivariate calibration framework to correct inter-component relationships, the proposed method effectively mitigates information loss during despeckling. Simultaneously, a global-local interaction strategy combined with edge-aware enhancement techniques is employed to comprehensively extract feature information from sonar images. To further enhance the effectiveness of the proposed method, a novel loss function is designed to account for image content loss, identity loss, and zero loss within the multivariate calibration process, ensuring more precise optimization and improving despeckling performance in sonar imaging. Extensive experiments on real-world sonar image datasets demonstrate that our method achieves superior despeckling results compared to state-of-the-art approaches.