STGAN: Sonar Image Despeckling Method Utilizing GAN and Transformer

This study presents an innovative method, denoted as STGAN, aimed at enhancing the quality of sonar images by addressing the pervasive issue of speckle noise. The proposed approach leverages a generative adversarial network framework, wherein Transformer blocks are harnessed to capture global features, while Convolution blocks focus on extracting local features within both the generator and discriminator architectures. Through adversarial training, STGAN acquires a holistic understanding of feature distributions from the training data, resulting in the production of high-fidelity denoised images. A novel loss function is introduced, amalgamating adversarial loss, content preservation, local texture style, and global similarity considerations. This multifaceted loss function effectively mitigates image distortion and information loss inherent in feature extraction. Empirical evaluations, conducted on synthetic speckle noise images and authentic sonar data, substantiate STGAN’s remarkable denoising capabilities, showcasing performance enhancements of approximately 10.70%–58.73% and 2.74%–22.97%, respectively, over established baseline methods.