Residual Contrastive Learning for Image Reconstruction: Learning Transferable Representations from Noisy Images

TL;DR

This paper introduces Residual Contrastive Learning (RCL), a novel unsupervised framework that leverages residuals for low-level image restoration, improving transferability and reducing annotation costs compared to existing methods.

Contribution

The paper proposes a residual contrastive learning paradigm that aligns contrastive pretext tasks with image reconstruction, enhancing robustness and transferability for noisy image tasks.

Findings

RCL outperforms recent self-supervised methods on denoising and super-resolution.

Unsupervised pre-training with RCL reduces annotation costs significantly.

RCL learns robust representations suitable for various low-level vision tasks.

Abstract

This paper is concerned with contrastive learning (CL) for low-level image restoration and enhancement tasks. We propose a new label-efficient learning paradigm based on residuals, residual contrastive learning (RCL), and derive an unsupervised visual representation learning framework, suitable for low-level vision tasks with noisy inputs. While supervised image reconstruction aims to minimize residual terms directly, RCL alternatively builds a connection between residuals and CL by defining a novel instance discrimination pretext task, using residuals as the discriminative feature. Our formulation mitigates the severe task misalignment between instance discrimination pretext tasks and downstream image reconstruction tasks, present in existing CL frameworks. Experimentally, we find that RCL can learn robust and transferable representations that improve the performance of various downstream tasks, such as denoising and super resolution, in comparison with recent self-supervised methods designed specifically for noisy inputs. Additionally, our unsupervised pre-training can significantly reduce annotation costs whilst maintaining performance competitive with fully-supervised image reconstruction.