Despite being widely used to lower the cost of image storage and transmission, JPEG compression introduces unwanted artifacts like ringing, blurring, and blocking, especially at high compression ratios. These artifacts seriously impair visual perception and make it more difficult to perform later image analysis tasks like enhancement and recognition. The research aim is to reduce computational complexity while effectively recovering high-quality images from JPEG-compressed inputs. This is done by incorporating discrete cosine transform (DCT) priors by a pixel position labeling map and quantization table into a lightweight convolutional neural network (CNN). The network applies a recursive pixel-shuffling mechanism to extend the receptive field and enhance feature learning with a three-level connection (residual) structure incorporating both global, local, and intermediate residuals. Tests on benchmark images (LIVE1, Classic5, and DIV2K) demonstrate that the model is more effective than the conventional methods of removing artifacts in JPEG against the background of computational efficiency, which are measured in terms of peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and perceptual quality of the proposed model.
Balaji et al. (Fri,) studied this question.