Convolutional Neural Network for Image Compression with Application to JPEG Standard

In this paper the authors propose a novel structure of convolutional neural network for lossy compression of images intended to be used as an extension of JPEG image compression standard. The convolutional network is trained on the set of images randomly selected from the database of high-quality images representing human faces and its effectiveness is verified experimentally using both human faces images as well as standard test images. The performance of the proposed network expressed in terms of its compression capabilities and image reconstruction quality is compared to other approaches utilizing the standard Discrete Cosine Transform, Lapped Orthogonal Transform, Modulated Lapped Transform and Karhunen-Loeve Transform, also incorporated into JPEG image compression standard. The obtained experimental results indicate that the proposed approach not only performs significantly better than the remaining approaches in terms of objective image quality measures, but also enables significant reduction of the blocking defects, which was verified by visual examination, when compared to the remaining tested transforms.