Towards Effective Multiple-in-One Image Restoration: A Sequential and Prompt Learning Strategy

TL;DR

The paper defines the challenging problem of training a single model to perform seven diverse image restoration tasks (MiO IR) and identifies two core hurdles: optimizing heterogeneous objectives and adapting to multiple tasks. It proposes sequential learning to stabilize cross-task optimization and prompt learning (explicit and adaptive) to enhance task awareness, showing consistent gains across CNN and Transformer backbones. Through extensive experiments on In-Dis, Out-Dis, and Unknown test sets with diverse backbones, the methods improve PSNR and can complement existing MiO IR approaches, even enabling restoration-style control via prompt interpolation. The work demonstrates improved generalization to unseen degradations, highlights the importance of task-aware degradation representations, and provides a practical framework for building more generalizable image restoration models.

Abstract

While single task image restoration (IR) has achieved significant successes, it remains a challenging issue to train a single model which can tackle multiple IR tasks. In this work, we investigate in-depth the multiple-in-one (MiO) IR problem, which comprises seven popular IR tasks. We point out that MiO IR faces two pivotal challenges: the optimization of diverse objectives and the adaptation to multiple tasks. To tackle these challenges, we present two simple yet effective strategies. The first strategy, referred to as sequential learning, attempts to address how to optimize the diverse objectives, which guides the network to incrementally learn individual IR tasks in a sequential manner rather than mixing them together. The second strategy, i.e., prompt learning, attempts to address how to adapt to the different IR tasks, which assists the network to understand the specific task and improves the generalization ability. By evaluating on 19 test sets, we demonstrate that the sequential and prompt learning strategies can significantly enhance the MiO performance of commonly used CNN and Transformer backbones. Our experiments also reveal that the two strategies can supplement each other to learn better degradation representations and enhance the model robustness. It is expected that our proposed MiO IR formulation and strategies could facilitate the research on how to train IR models with higher generalization capabilities.

Figures (8)

• Figure 1: Our proposed sequential and prompt learning strategies could improve the performance of both CNN and Transformer backbones on in/out-of-distribution test sets. '-M' refers to mixed learning. '-S+EP' or '-S+AP' refers to using both sequential learning and explicit or adaptive prompt learning.
• Figure 2: (a) Overview of the MiO IR problem, which has 7 IR tasks. (b) The proposed sequential learning strategy. (c) The proposed prompt learning strategy. We provide two specific methods, explicit prompt learning and adaptive prompt learning.
• Figure 3: Visual comparison of the results of models on the 7 In-Dis MiO test sets. (Zoom in and follow the arrows for the best view).
• Figure 4: The clusters of the prompt feature. We use the features after $F_{ext}(P)$ to analyze the degradation representation. A higher CHI indicates a stronger clustering performance.
• Figure 5: Image restoration style adjustment of Restormer-S+EP by interpolating the prompts of Low-light enhancement and Deraining.