EchoIR: Advancing Image Restoration with Echo Upsampling and Bi-Level Optimization

TL;DR

EchoIR addresses upsampling-induced information loss in image restoration by introducing a bilaterally guided Echo-Upsampler and a training framework AS-BLO that converts a bi-level optimization into sequential single-level problems solvable by gradient descent. The architecture combines shallow feature extraction, Mix-Attention blocks, and encoder-guided upsampling to deliver high-fidelity restoration across deraining, deblurring, and denoising. AS-BLO provides a principled, convergent training regime that jointly optimizes restoration and upsampling components, yielding state-of-the-art results on standard benchmarks. Overall, EchoIR demonstrates that leveraging bilateral encoder echoes and tractable bi-level optimization can substantially improve the quality of restored images in practical low-level vision tasks.

Abstract

Image restoration represents a fundamental challenge in low-level vision, focusing on reconstructing high-quality images from their degraded counterparts. With the rapid advancement of deep learning technologies, transformer-based methods with pyramid structures have advanced the field by capturing long-range cross-scale spatial interaction. Despite its popularity, the degradation of essential features during the upsampling process notably compromised the restoration performance, resulting in suboptimal reconstruction outcomes. We introduce the EchoIR, an UNet-like image restoration network with a bilateral learnable upsampling mechanism to bridge this gap. Specifically, we proposed the Echo-Upsampler that optimizes the upsampling process by learning from the bilateral intermediate features of U-Net, the "Echo", aiming for a more refined restoration by minimizing the degradation during upsampling. In pursuit of modeling a hierarchical model of image restoration and upsampling tasks, we propose the Approximated Sequential Bi-level Optimization (AS-BLO), an advanced bi-level optimization model establishing a relationship between upsampling learning and image restoration tasks. Extensive experiments against the state-of-the-art (SOTA) methods demonstrate the proposed EchoIR surpasses the existing methods, achieving SOTA performance in image restoration tasks.

Figures (5)

• Figure 1: Schematic diagram of the Echo-Upsampler and comparison of the EchoIR model with other methods in deraining and denoising tasks. MAM is Mix-attention module.
• Figure 2: The structure of EchoIR. SD is the simple downsampler. EU is Echo-upsampler. LN is layer normalization, SA is multi-head self-attention, CA is channel attention, and GDFN is gated-Dconv FN.
• Figure 3: Visual comparison of results with SOTA draining methods. This comparison employs the output results for direct comparison, thereby providing a clear reflection of the quality of the model's output outcomes.
• Figure 4: Visual comparison of the result with SOTA deblurring methods. This comparison employs the output results for direct comparison, thereby providing a clear reflection of the quality of the model's output outcomes.
• Figure 5: Visual comparison of the result with SOTA denosing methods. This comparison employs the difference map for clear comparison, thereby providing a clear reflection of the quality of the model's output outcomes.