GRIDS: Grouped Multiple-Degradation Restoration with Image Degradation Similarity

TL;DR

GRIDS tackles multi-degradation image restoration by quantifying relationships between degradations through deep degradation representations and a Generalized Gaussian Distribution-based KL-divergence similarity, forming a degradation similarity matrix. It then solves a grouping optimization to partition 11 degradations into a minimal number of highly correlated groups, training a dedicated mix-training model for each group under a performance-drop constraint $\Delta P(C_k) \leq \delta$. At inference, GRIDS automatically selects the most appropriate group via an adaptive model selection mechanism, and it can predict generalization ability without running inference by comparing degradation distributions. Empirical results on 11 degradations show GRIDS achieves up to $2.24$ dB improvement over a mix-training baseline and competitive performance with single-task upper bounds, while enabling efficient handling of unknown real-world degradations.

Abstract

Traditional single-task image restoration methods excel in handling specific degradation types but struggle with multiple degradations. To address this limitation, we propose Grouped Restoration with Image Degradation Similarity (GRIDS), a novel approach that harmonizes the competing objectives inherent in multiple-degradation restoration. We first introduce a quantitative method for assessing relationships between image degradations using statistical modeling of deep degradation representations. This analysis facilitates the strategic grouping of similar tasks, enhancing both the efficiency and effectiveness of the restoration process. Based on the degradation similarity, GRIDS divides restoration tasks into one of the optimal groups, where tasks within the same group are highly correlated. For instance, GRIDS effectively groups 11 degradation types into 4 cohesive groups. Trained models within each group show significant improvements, with an average improvement of 0.09dB over single-task upper bound models and 2.24dB over the mix-training baseline model. GRIDS incorporates an adaptive model selection mechanism for inference, automatically selecting the appropriate grouped-training model based on the input degradation. This mechanism is particularly useful for real-world scenarios with unknown degradations as it does not rely on explicit degradation classification modules. Furthermore, our method can predict model generalization ability without the need for network inference, providing valuable insights for practitioners.

Figures (6)

• Figure 1: Schematic comparison of our GRIDS with the previous all-in-one methods.
• Figure 2: When tasks are negatively correlated, optimizing them simultaneously often results in a significant performance decline across all tasks. Conversely, training closely related tasks together does not lead to severe performance deterioration.
• Figure 3: The obtained degradation similarity matrix $S$, which quantitatively describes the similarity relationship between 11 different degradations.
• Figure 4: Workflow of the similarity-based degradation grouping method. Initially, a pretrained degradation feature extractor is employed to extract deep degradation representations, which are then modeled using the Generalized Gaussian Distribution. By computing the KL divergence between these representations, a similarity matrix is generated. Subsequently, the degradation grouping search algorithm utilizes the similarity matrix to automatically classify the degradations into relevant groups.
• Figure 5: Thanks to the statistical modeling of degradations, GRIDS can automatically switch the optimal model for an unknown input image without any auxiliary classification module. By computing the KL divergence between the input image and predefined degradation groups, it can identify the most similar group for restoration and predicts model generalization ability and processing performance without actual inference.