LoC-LIC: Low Complexity Learned Image Coding Using Hierarchical Feature Transforms

TL;DR

This work tackles the high computational cost of learned image compression by introducing a hierarchical feature transform that assigns few channels to high-resolution feature maps and more channels to smaller, deeper maps. The approach yields a large reduction in forward MACs from about 1256 kMAC/Pixel to 270 kMAC/Pixel while preserving rate-distortion performance via a hyper-encoder/decoder with a multi-reference entropy model. Evaluations on standard benchmarks show competitive RD performance with substantially lower complexity, though training on 256x256 crops can limit performance on large images. The method enables efficient deployment on diverse devices and opens avenues for new architecture designs in learned image compression.

Abstract

Current learned image compression models typically exhibit high complexity, which demands significant computational resources. To overcome these challenges, we propose an innovative approach that employs hierarchical feature extraction transforms to significantly reduce complexity while preserving bit rate reduction efficiency. Our novel architecture achieves this by using fewer channels for high spatial resolution inputs/feature maps. On the other hand, feature maps with a large number of channels have reduced spatial dimensions, thereby cutting down on computational load without sacrificing performance. This strategy effectively reduces the forward pass complexity from \(1256 \, \text{kMAC/Pixel}\) to just \(270 \, \text{kMAC/Pixel}\). As a result, the reduced complexity model can open the way for learned image compression models to operate efficiently across various devices and pave the way for the development of new architectures in image compression technology.

Figures (6)

• Figure 1: Overall view of the proposed architecture with hierarchical feature encoder and decoder.
• Figure 2: The compression efficiency vs complexity of different learned image compression models. The complexity is measured in terms of (kMAC/Pixel). (a) BD-PSNR (b) BD-SSIM.
• Figure 3: Comparison of forward complexity between our approach and various learned image compression models
• Figure 4: Assessments and comparisons of image compression models using different metrics and datasets. (a) PSNR scores on the Kodak dataset, (b) MS-SSIM scores on the Kodak dataset, and (c) PSNR scores on the CLIC Professional Valid 2020 dataset.
• Figure 5: Our novel approach performance compared to MLIC++ and LIC-TCM on image num. 3 from the CLIC Professional Valid 2020 dataset.