Self-Supervised Denoising Method for Sonar Images Based on Spatio-temporal Correlation

In complex underwater scenarios, sonar imaging serves as a fundamental and indispensable technique for marine exploration and underwater environmental perception. However, raw sonar images inherently suffer from low contrast, limited spatial resolution, and severe speckle noise. These inherent degradation characteristics severely deteriorate imaging quality and inevitably hinder reliable visual interpretation as well as the implementation of downstream high-level visual analysis tasks. Although existing mainstream denoising algorithms can suppress partial noise interference to a certain extent, they typically induce over-smoothing artifacts and irreversible loss of fine texture features, which substantially degrade the practical applicability of these methods in real underwater scenarios. To mitigate the aforementioned limitations, this paper proposes a self-supervised spatio-temporal denoising framework tailored for sequential sonar images. The proposed method constructs effective self-supervised training pairs by fully exploiting the inherent spatio-temporal correlation among consecutive sonar frames. Furthermore, an adversarial loss function is integrated into the framework to preserve detailed texture information and suppress excessive blurring distortion during the denoising process. Extensive experiments conducted on real-world sonar datasets with diverse scene conditions validate that the proposed method outperforms conventional denoising approaches in terms of visual restoration performance and multiple authoritative no-reference image quality metrics. In addition, the proposed method significantly enhances the detection precision of downstream object detection models under complex noisy underwater environments. This work effectively alleviates the performance gap and establishes a robust functional linkage between low-level sonar image enhancement and high-level intelligent visual perception tasks.