Powerful Lossy Compression for Noisy Images

TL;DR

This paper designs an end-to-end trainable network, which includes the main encoder branch, the guidance branch, and the signal-to-noise ratio (SNR) aware branch, and demonstrates that the joint solution outperforms existing state-of-the-art methods.

Abstract

Image compression and denoising represent fundamental challenges in image processing with many real-world applications. To address practical demands, current solutions can be categorized into two main strategies: 1) sequential method; and 2) joint method. However, sequential methods have the disadvantage of error accumulation as there is information loss between multiple individual models. Recently, the academic community began to make some attempts to tackle this problem through end-to-end joint methods. Most of them ignore that different regions of noisy images have different characteristics. To solve these problems, in this paper, our proposed signal-to-noise ratio~(SNR) aware joint solution exploits local and non-local features for image compression and denoising simultaneously. We design an end-to-end trainable network, which includes the main encoder branch, the guidance branch, and the signal-to-noise ratio~(SNR) aware branch. We conducted extensive experiments on both synthetic and real-world datasets, demonstrating that our joint solution outperforms existing state-of-the-art methods.

Figures (11)

• Figure 1: The network architecture of proposed SNR-aware joint solution of image compression and denoising. The architecture contains three branches, "Teacher Guidance Branch", "Main Encoder Branch" and "SNR Aware Branch", in the left half of the figure. The right half of the figure contains the main decoder, entropy models, context model, and hyper encoder/decoder commonly used in recent learning-based compression method cheng2020learned. Note that the "Teacher Guidance Branch" is for training only, the "Main Encoder Branch" and "SNR Aware Branch" are activated during training of the entire network and used for inference. $\bigoplus$ denotes the addition by element.
• Figure 2: Overall RD curves for the Kodak and CLIC datasets across all noise levels. Our proposed joint solution, indicated by red curves, exhibits superior RD performance compared to pure compression, sequential, and joint methods.
• Figure 3: RD curves on the Kodak dataset at various noise levels. Our method surpasses both sequential and joint methods, particularly at the high noise level. This denotes that the proposed SNR-aware branch efficiently captures valuable information through a combination of local and non-local features.
• Figure 4: RD performance curves optimized by MSE aggregated on SIDD. Our proposed method achieved the best RD performance. This indicates that our method is robust on real noisy images. The purple dotted line serves as a reference for the DeamNet ren2021adaptive ideal case without compression.
• Figure 5: RD curves on the CLIC dataset at various noise levels. Our method surpasses the state-of-the-art joint method, particularly at the high noise level. This denotes that the proposed SNR-aware branch efficiently captures valuable information through a combination of local and non-local features.