Exploration of Learned Lifting-Based Transform Structures for Fully Scalable and Accessible Wavelet-Like Image Compression (2402.18761v1)

Abstract: 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.

• The paper introduces novel learned lifting structures that merge neural networks with wavelet-like transforms to enhance image compression scalability.
• It demonstrates that a hybrid structure with additional learned lifting steps and the proposal-opacity topology significantly boosts coding efficiency without incurring excessive computational cost.
• The study outlines an optimal balance between improved performance and increased complexity, offering guidance for future research in scalable image compression techniques.

Exploration of Learned Wavelet-Like Transforms for Enhanced Image Compression

Introducing New Transform Structures

The evolution of learning-based methods has transformed the landscape of image compression. Among these, lifting-based, wavelet-like transforms have emerged as a pivotal area of research due to their inherent multi-resolution support and scalability features. Notably, this paper offers an in-depth analysis of integrating neural networks into lifting-based transforms for scalable image compression, marking a significant stride in the quest to refine wavelet-based methods with the power of machine learning.

Investigation of Lifting Structures

The research primarily revolves around the exploration of different ways to incorporate learned lifting steps within wavelet-like transforms and their impact on compression efficiency. Among various configurations, the paper prominently features:

• Predict-update and update-predict lifting structures, where conventional lifting operators are substituted with learned neural networks.
• The introduction of a hybrid lifting structure that augments a base wavelet transform with additional learned lifting steps for aliasing suppression and redundancy reduction.

The comparison among these structures yielded insightful conclusions. Notably, the hybrid structure with two additional learned lifting steps, when employing the proposal-opacity network topology, showcased superior performance.

The Proposal-Opacity Network Topology

A distinguished contribution of this paper is the proposal-opacity network topology. Characterized by linear proposals modulated by non-linear opacities, this topology offers a nuanced approach to compressing images. It demonstrates how increasing the diversity, denoted by the number of channels within each learned lifting operator, can significantly enhance coding efficiency. Impressively, the research revealed that employing a more considerable number of channels, within reasonable bounds, contributed more effectively to compression performance than simply increasing the depth of lifting structures or the spatial support.

Computational Considerations and Practical Implications

On the computational front, the paper meticulously evaluates the trade-offs between the enhanced coding efficiency afforded by learned lifting operators and the associated increase in computational complexity and region of support. The findings advocate for a balanced approach, recommending the augmentation of well-performing base wavelet transforms with a selective number of learned lifting steps to achieve an optimal blend of compression efficiency, computational load, and support compactness.

Future Directions

Looking forward, this comprehensive examination of learned lifting-based transform structures opens several avenues for further research. The demonstrated superiority of the proposal-opacity network topology paves the way for its application in other domains of image processing and beyond. Moreover, the paper underlines the importance of continuing to probe into the balance between network complexity and compression efficiency—a key consideration as the field progresses.

Conclusions

In summary, this paper stands as a rigorous exploration of leveraging neural networks within the framework of lifting-based wavelet-like transforms for scalable image compression. Its findings not only underscore the potential of learned lifting operators in enhancing coding efficiency but also highlight the critical role of choosing the appropriate network topologies and lifting structures. As the pursuit of more advanced and efficient image compression technologies marches on, the insights from this research offer valuable guidance for future explorations in the field of generative AI and beyond.