TL;DR
This paper introduces AIBNet, an adaptive image deblurring network that identifies and restores blurred regions differently, utilizing novel modules for high-frequency feature selection and a progressive training strategy, leading to superior results.
Contribution
The paper presents a novel AIBNet architecture with adaptive blurred region identification and specialized modules, improving deblurring accuracy over existing methods.
Findings
AIBNet outperforms state-of-the-art deblurring methods in experiments.
The spatial feature differential handling enhances focus on blurred regions.
High-frequency feature selection improves restoration of fine details.
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…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
1. A spatial feature differential handling block is introduced to enable the model to focus on key information in the blurred regions. 2. A high-frequency feature section block is proposed to retain the most important high-frequency regions. 3. Technically, a progressive training strategy is used to save GPU memory and leads to performance improvements. 4. The model achieves promising performance on both synthetic and real-world datasets.
1. The claim `most overlook the fact that the degradation degree varies across different regions` appears inappropriate. This topic has been extensively studied in recent years and is commonly referred to as spatially variant degradation [1]. 2. Several recent and important references are missing from the comparative analysis, such as the Mamba-based deblurring method EVSSM [2]. In addition, compared with EVSSM, the proposed model in this paper involves a larger number of parameters. 3. The no
1. The paper shows good results on the GoPro, HIDE and RealBlur datasets. 2. Progressive decoder reduced the training burden
1. Please write T to the top right corer of Q. It is misleading in Eqn. 2 and 4. 2. Performance heavily relies on a pre-trained encoder. 3. The mathematical analysis of SFEM is not enough.
AIBNet introduces a novel architecture that combines spatial-domain differencing and frequency-domain feature selection to distinguish blurred regions. The SFEM integrates the principle of a differential amplifier into visual feature modeling, offering theoretical insight. The proposed model achieves superior performance across multiple datasets (e.g., GoPro and HIDE), while the progressive training strategy reduces computational and memory overhead.
However, a major weakness of this work lies in the discrepancy between its stated goal and experimental evidence. The title and abstract emphasize “Adaptive Identification of Blurred Regions”, yet the authors provide no direct proof that their method can indeed distinguish blurred regions. For example, there are no visualizations or heatmaps illustrating which parts of an image are identified as blurred versus sharp. This omission significantly undermines the paper’s core claim and overall persu
1. The motivation and the proposed modules, including SFDHBlock and HFSBlock, are well-reasoned and appropriate for the image deblurring task. 2. The proposed architecture is elegant and easy to follow. 3. The paper is well written and clearly presented.
1. This work appears to be primarily an engineering effort that lacks sufficient novelty in the field of image deblurring. The proposed modules, including the use of a pre-trained encoder, SFDHBlock, and HFSBlock, are relatively straightforward ideas aimed at improving performance. 2. The proposed SFEM, which subtracts two attention maps, is quite similar to the "Differential Transformer" published at ICLR 2025, which reduces the novelty and contribution of SFEM. 3. Since the overall architect
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
