Adaptive Identification of Blurred Regions for Accurate Image Deblurring

TL;DR

This work addresses image deblurring when blur severity varies across regions by introducing AIBNet, which adaptively identifies blurred regions in both spatial and frequency domains. The decoder-centric design employs SFDHBlock (featuring SFEM and SCA) to emphasize blurred areas and suppress non-blurred information, along with HFSBlock for selective high-frequency feature retention, all powered by a frozen pre-trained encoder. A progressive training strategy trains sub-decoders sequentially to reduce memory demands, while a composite loss balances pixel, edge, and frequency-domain cues. Empirically, AIBNet achieves state-of-the-art performance on synthetic (GoPro) and real-world (RealBlur) datasets and demonstrates strong generalization and resource-efficiency through ablations and scalability analyses.

Abstract

Image deblurring aims to restore high-quality images from blurred ones. While existing deblurring methods have made significant progress, most overlook the fact that the degradation degree varies across different regions. In this paper, we propose AIBNet, a network that adaptively identifies the blurred regions, enabling differential restoration of these regions. Specifically, we design a spatial feature differential handling block (SFDHBlock), with the core being the spatial domain feature enhancement module (SFEM). Through the feature difference operation, SFEM not only helps the model focus on the key information in the blurred regions but also eliminates the interference of implicit noise. Additionally, based on the fact that the difference between sharp and blurred images primarily lies in the high-frequency components, we propose a high-frequency feature selection block (HFSBlock). The HFSBlock first uses learnable filters to extract high-frequency features and then selectively retains the most important ones. To fully leverage the decoder's potential, we use a pre-trained model as the encoder and incorporate the above modules only in the decoder. Finally, to alleviate the resource burden during training, we introduce a progressive training strategy. Extensive experiments demonstrate that our AIBNet achieves superior performance in image deblurring.

Figures (6)

• Figure 1: Varying degrees of degradation across different regions. (a) is the blurred image, (b) is the corresponding clear image, and (c) is the residual image between the blurred and clear images.
• Figure 2: (a) The overall architecture of the proposed AIBNet. (b) The decoder, consisting of N spatial feature differential handling blocks (SFDHBlocks) and a high-frequency feature selection block (HFSBlock). (c) The structure of the HFSBlock is shown, with the case of using a single mask matrix for simplicity. (d) The SFDHBlock, which consists of two branches: the SCA proposed in NAFNet chen2022simple and the spatial feature enhancement module (SFEM). (e) The structure of the SFEM.
• Figure 3: Image deblurring comparisons on the synthetic dataset Gopro.
• Figure 4: Image deblurring comparisons on the real-world dataset realblurrim_2020_ECCV.
• Figure 5: The internal features of the SFEM. With our spatial feature enhancement, the SFEM captures finer details, such as the number plate, compared to the initial features. Zoom in for a clearer view.