An Image Compression Framework Based on Multi-scale Convolutional Neural Network for Deformation Images

Along with the development of deep learning, convolutional neural networks (CNN) have become more and more widely used in the field of image compression, and have made image compression technology significantly improved on performance and cost. In order to make the image better preserve the details and texture of the original image while compressing, we propose a novel method to optimize image compression. We first focus on slight deformation image which is changed by original image, which means that we preserve the features of image detail information when the bits are not increased, and then the deformed image is transmitted to our network framework to achieve image compression and the reconstruction process. In our system, first, a multi-scale convolutional neural network is used to learn the best compact representation from the input image to achieve the purpose of extracting multiscale structural information of natural images. The result of compact representation is then encoded and decoded by a conventional image codec. Finally, a reconstructed convolutional neural network is used for reconstructing the decoded image with high quality and accurate quality at the decoding end. Experimental results demonstrate that our network is superior to most existing methods and can improve image quality with more visual detail.