Research on Denoising of Multiple Compressed Image Based on Structure-Texture Decomposition

Abstract

At present, most of the researches on denoising of compressed images are one-time compression. But in practice, the image will be compressed more than once. Therefore, a denoising method combining deep learning and traditional methods was proposed for multiple compressed images. First, the data set was compressed twice through singular value decomposition (SVD) to obtain noisy images after multiple compressions. Secondly, the noisy image was decomposed to obtain the noisy structure image and texture image. Then denoised them separately, the noisy structure image was used the feed-forward denoising convolutional neural network (DnCNN), and the noisy texture image was used the selected mean method. Finally, the denoised structure image and texture image were combined to obtain the denoised multiple compressed image. The Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) of this method are approximately improved by 0.9~1.5dB and 0.02~0.06, respectively. Moreover, the texture image is extracted and targeted for denoising, retaining more detailed information and achieving a clearer visual effect.