This paper provides a comprehensive study on features and performance of different ways to incorporate neural networks into lifting-based wavelet-like transforms, within the context of fully scalable and accessible image compression. Specifically, we explore different arrangements of lifting steps, as well as various network architectures for learned lifting operators. Moreover, we examine the impact of the number of learned lifting steps, the number of channels, the number of layers and the support of kernels in each learned lifting operator. To facilitate the study, we investigate two generic training methodologies that are simultaneously appropriate to a wide variety of lifting structures considered. Experimental results ultimately suggest that retaining fixed lifting steps from the base wavelet transform is highly beneficial. Moreover, we demonstrate that employing more learned lifting steps and more layers in each learned lifting operator do not contribute strongly to the compression performance. However, benefits can be obtained by utilizing more channels in each learned lifting operator. Ultimately, the learned wavelet-like transform proposed in this paper achieves over 25% bit-rate savings compared to JPEG 2000 with compact spatial support.

This paper studies features and performance of different structures for learned wavelet-like transforms in fully scalable image compression. Specifically, we explore different neural network topologies and various arrangements of lifting steps to improve the existing wavelet transform. Experimental results strongly suggest that the proposed proposal-opacity network topology, comprising a collection of linear predictions modulated by non-linear opacities, performs better than the other considered designs. Results also show that augmenting a good base wavelet transform with two learned lifting steps performs significantly better than other learned lifting structures, achieving 26.8% bit-rate savings over JPEG 2000.

This paper proposes a secondary transform to improve wavelet-based image compression schemes such as JPEG 2000. The proposed approach includes two neural network steps, a high-to-low step followed by a low-to-high step. The high-to-low step suppresses aliasing in the low-pass band by using the detail bands at the same resolution, while the low-to-high step aims to further remove redundant information from the detail bands so as to achieve higher energy compaction. We employ the same neural networks for these two steps at every level of the hierarchical discrete wavelet transform (DWT) decomposition. We demonstrate that the visual quality of the LL bands at different resolutions can be dramatically enhanced, along with greatly improved coding efficiency up to 2dB compared with JPEG 2000, while preserving the full quality and resolution scalability and spatial random access features of JPEG 2000.

This paper proposes a convolutional neural network (CNN) based secondary transform for not only improved coding efficiency in the JPEG2000 image compression format, but also to produce more appealing approximation sub-bands at different resolutions. The CNN in this work exploits information in detail sub-bands to predict some of the aliasing information in the corresponding approximation or low-pass sub-band; this reduction in aliasing, although not perfect, improves the compressibility of the “cleaned” approximation subband. This process is repeated in subsequent wavelet decomposition levels to further improve coding efficiency. Experimental results show that, at high bit rates, the proposed network outperforms conventional JPEG2000 compression framework by up to 1.2 dB, especially for images with strong geometric flow.

This paper proposes a secondary transform for wavelet based image compression, whose aim is to predict detail sub-bands from the corresponding low-pass sub-band so as to improve energy compaction. The proposed prediction strategy exploits the orientation of local image features to untangle aliasing components in the low-pass sub-band, and achieves prediction by projecting the cleaned synthesized low-pass sub-band back into the wavelet domain. At low bit rates, the proposed scheme can considerably reduce distortion in a JPEG 2000 based compression framework, especially for images with strong geometric flow.