STMLN: Transformer-Based Multivariate Loop Recurrent Denoising Network for Sonar Image

The impact of speckle noise on sonar image quality during the imaging process is particularly pronounced. To enhance the quality of sonar images, this paper explores a speckle noise reduction network for sonar images based on Transformer and multivariate loop operations, namely STMLN. This network architecture integrates the Transformer into the denoising network with the aim of extracting global features and combining them with convolutional operations for a more comprehensive understanding of sonar image characteristics. Unlike conventional models, STMLN decomposes model outputs into denoised images, signal-related noise, and signal-independent noise, refining the output through multivariate loop operations to enhance denoising accuracy. Furthermore, we designed a specific loss function addressing the correlation between sonar image content and multivariate loop operations to calibrate the model's learning process, reducing information loss and enhancing output image quality. Experimental results on synthetic speckle noise datasets and real sonar image datasets demonstrate that compared to existing sonar image denoising methods, STMLN achieves outstanding denoising performance.