An Image Compression Processing Method Based On Deep Learning

In recent years, the combination of neural network and deep learning has made outstanding achievements in image compression, recognition and other fields. However, the study found that when using a given generative adversarial network (GAN) compression model, the color image compression performance is better than the gray image. In order to improve the compression effect of gray image, a post-processing method is proposed in this paper. The main idea is to process each component of the image generated by the compression model, so as to reduce the influence of different weights of training parameters on the generated image. The method is divided into two cases. One is equal weight processing of component coefficients, which means that the generated component coefficients are all 1/3. The peak signal to noise ratio (PSNR) of images can be improved within the range of 0.10-0.41dB, and the average can be improved by 0.21dB. The other is the unequal weight processing of component coefficients. After data fitting, it is concluded that when the component coefficients are 0.08, 0.49 and 0.35 respectively, the PSNR can be increased within the range of 0.15-1.1dB, and the average can be increased by 0.55dB. Therefore, this method can effectively improve the compression quality of gray image.