X-DECODE: EXtreme Deblurring with Curriculum Optimization and Domain Equalization

TL;DR

X-DECODE tackles extreme image deblurring by employing curriculum learning to gradually expose a GAN-based deblurring model to increasing blur severity, coupled with a hybrid loss that combines perceptual, L1, and hinge components. A linear curriculum emerges as the strongest progression, and introducing a small fraction of sharp-sharp pairs enhances generalization across blur levels and domains, achieving notable gains on Extreme-GoPro and Extreme-KITTI. The framework is architecture-agnostic and demonstrates robust cross-domain performance, highlighting the value of structured difficulty progression for extreme artifact restoration. These findings suggest practical pathways for extending curriculum-based domain management to other extreme artifacts and for improving real-world applicability of deblurring systems.

Abstract

Restoring severely blurred images remains a significant challenge in computer vision, impacting applications in autonomous driving, medical imaging, and photography. This paper introduces a novel training strategy based on curriculum learning to improve the robustness of deep learning models for extreme image deblurring. Unlike conventional approaches that train on only low to moderate blur levels, our method progressively increases the difficulty by introducing images with higher blur severity over time, allowing the model to adapt incrementally. Additionally, we integrate perceptual and hinge loss during training to enhance fine detail restoration and improve training stability. We experimented with various curriculum learning strategies and explored the impact of the train-test domain gap on the deblurring performance. Experimental results on the Extreme-GoPro dataset showed that our method outperforms the next best method by 14% in SSIM, whereas experiments on the Extreme-KITTI dataset showed that our method outperforms the next best by 18% in SSIM. Ablation studies showed that a linear curriculum progression outperforms step-wise, sigmoid, and exponential progressions, while hyperparameter settings such as the training blur percentage and loss function formulation all play important roles in addressing extreme blur artifacts. Datasets and code are available at https://github.com/RAPTOR-MSSTATE/XDECODE

Figures (9)

• Figure 1: Current deblurring methods achieve effective results only for low-to-moderate blur levels but struggle with extreme blur cases. Our approach bridges this gap by using curriculum learning to train a model to handle extreme degradations.
• Figure 2: Proposed training strategy for X-DECODE with controlled varying of blur levels
• Figure 3: Illustration of a general curriculum strategy for training a deblurring model, where the task difficulty starts with an initial blur level of 3 and increases at every epoch $i$ until the maximum blur level $X$ is reached at $N$ epochs.
• Figure 4: Different blur level progressions for curriculum learning (from left to right, top to bottom: step-wise, slower step-wise, linear, sigmoid, exponential)
• Figure 5: Blur variations in the Extreme-GoPro dataset. Columns show sharp images followed by BL 19, BL 25, and BL 29.