Continuous Patch Stitching for Block-wise Image Compression

TL;DR

The paper tackles block artefacts and high memory costs in learned block-wise image compression by introducing Continuous Patch Stitching (CPS), a padding-free framework that processes image patches in parallel while guaranteeing full-image equivalence. CPS relies on the Parallel Overlapping Patch Stitching (POPS) strategy to mathematically bound overlaps and ensure continuity, aided by padding-free residual blocks with even-sized kernels (EDRB/EURB) and a bottleneck residual block (BRB) to deepen the network without size mismatches. An entropy-model-based compression framework with joint autoregressive and hierarchical priors underpins the rate-distortion optimization, achieving state-of-the-art performance with significantly reduced memory, including successful ultra-high-resolution compression under 4GB. The approach yields complete elimination of block artefacts without post-processing and offers practical benefits for scalable distributed vision tasks and high-resolution media compression.

Abstract

Most recently, learned image compression methods have outpaced traditional hand-crafted standard codecs. However, their inference typically requires to input the whole image at the cost of heavy computing resources, especially for high-resolution image compression; otherwise, the block artefact can exist when compressed by blocks within existing learned image compression methods. To address this issue, we propose a novel continuous patch stitching (CPS) framework for block-wise image compression that is able to achieve seamlessly patch stitching and mathematically eliminate block artefact, thus capable of significantly reducing the required computing resources when compressing images. More specifically, the proposed CPS framework is achieved by padding-free operations throughout, with a newly established parallel overlapping stitching strategy to provide a general upper bound for ensuring the continuity. Upon this, we further propose functional residual blocks with even-sized kernels to achieve down-sampling and up-sampling, together with bottleneck residual blocks retaining feature size to increase network depth. Experimental results demonstrate that our CPS framework achieves the state-of-the-art performance against existing baselines, whilst requiring less than half of computing resources of existing models. Our code shall be released upon acceptance.

Figures (8)

• Figure 1: Illustration of the proposed CPS framework, which is able to efficiently compress image patches in parallel, with the mathematical equivalence to compressing a whole image. This also ensures the intrinsic continuity when stitching patches together to an image, even without training our network.
• Figure 2: Illustration of our overlapping patch stitching, in which the overlapping size $o$ provides the upper bound of step size to ensure the intrinsic continuity of two consecutive patches, when compressed by our CPS framework.
• Figure 3: Basic building blocks of our CPS framework. The arrows indicate up-sampling/down-sampling with a scale of 2. Tconv represents transposed convolution. We also illustrate the impacted areas after important operations, whereby the yellow blocks indicate areas affected by both patches.
• Figure 4: The overall architecture of the proposed network in our CPS framework for block-wise image compression.
• Figure 5: R-D performance covering a wide range of traditional and learned compression methods on Kodak and DIV8k dataset for PSNR.