Speckle noise is a common artifact in coherent imaging systems such as radar, ultrasound, and optical coherence tomography (OCT), leading to a degradation in image quality. Existing denoising methods often compromise image details along with unwanted artifacts. This study proposes four novel denoising architectures: Guided DAE, Attention Guided DAE, Guided DAES, and Attention Guided DAES. Compared to existing denoising autoencoder models (DAE, DAES) and other deep learning models, the Attention Guided DAES achieves superior performance, effectively removing speckle noise while preserving image features. It outperforms conventional DAEs, achieving a PSNR of 29.77 and SSIM of 0.7846, resulting in clearer images. This approach can be applied across various imaging modalities. The Attention Guided DAES paves the way for improved image analysis in other domains, scientific research, and other fields by overcoming limitations imposed by speckle noise.
Su-ada et al. (Fri,) studied this question.