In this paper the authors present a novel structure of convolutional neural network for lossy image compression intended for use as a part of JPEG’s standard image compression stream. The network is trained on randomly selected images from high-quality image dataset of human faces and its effectiveness is verified experimentally using standard test images. Performance of the proposed network in terms of its compression capabilities and image reconstruction quality is compared with other approaches utilizing the standard Discrete Cosine Transform, Lapped Orthogonal Transform, Modulated Lapped Transform and Karhunen-Loeve Transform, embedded in the JPEG’s image compression stream. Experimental results indicate that the proposed solution not only performs significantly better than the remaining approaches in terms of compression capabilities and objective image quality measures, but also enables significant reduction of the tiling effect, which is noticeably present in the images processed with the remaining tested transforms.