De-compressed noise method based on image perception quality

Abstract

The invention discloses a compressed noise removing method based on image perception quality. Aiming at the characteristics of diversity, instability and the like of noise caused by video compression,a data set is constructed, and a deep residual denoising model is provided; meanwhile, in order to solve the problem that high-frequency information is lost in the denoising process of an existing algorithm, loss calculation based on visual perception characteristics is provided, a denoising model is assisted in learning noise residual errors, and the deep learning technology is utilized to learnactual compression noise to obtain a denoised image. The beneficial effects of the invention are that the method achieves the learning of the actual compression noise through the deep learning technology, and obtains more accurate residual noise; characterization image perception quality characteristics are extracted by using an image quality evaluation model and are used for loss calculation ofa denoising model, so that the denoised image better conforms to visual perception of people.

Description

technical field [0001] The present invention relates to the technical field of image processing, in particular to a method for decompressing noise based on image perception quality. Background technique [0002] In real life, the process of image digitization and transmission is often affected by imaging equipment and external environmental noise interference, which leads to a decrease in the quality of the acquired image. However, in many image application fields, the requirements for image quality are very high. For example, the image quality of video decreases after compression, and factors such as compression method, compressed bit rate, and video content will cause different noises in the image. Many current denoising methods target a specific noise type, such as Gaussian noise. However, the actual noise is diverse, and there may be multiple noises on an image, which poses a huge challenge to image denoising. [0003] In recent years, with the continuous development ...

AI Technical Summary

Problems solved by technology

At present, most denoising models based on deep learning use mean square error (MSE) as the loss function of the network, but MSE calculates the average value of pixel-level errors. During the training process, the model continuously learns, that is, minimizes MSE value, although the image generated in this way can obtain a high peak signal-to-noise ratio (PSNR), it often loses high-frequency information, and the image is too smooth, giving people a poor visual perception

Images