ExpRDiff: Short-exposure Guided Diffusion Model for Realistic Local Motion Deblurring

TL;DR

ExpRDiff tackles local motion blur in mobile photography by fusing context-aware blur detection, short-exposure guided restoration, and diffusion-based priors. It introduces a context-based local blur detector to identify blurry regions, a blurry-aware restoration that selectively uses short-exposure guidance, and a short-exposure guided diffusion model with ExpBFusion to produce realistic restorations. The method demonstrates superior performance on synthetic and real local-blur datasets, with ablations confirming the contribution of each component. This approach offers a practical path to high-quality deblurring in real-world, smartphone-driven imaging pipelines, potentially improving visual fidelity in dynamic scenes with limited lighting.

Abstract

Removing blur caused by moving objects is challenging, as the moving objects are usually significantly blurry while the static background remains clear. Existing methods that rely on local blur detection often suffer from inaccuracies and cannot generate satisfactory results when focusing solely on blurred regions. To overcome these problems, we first design a context-based local blur detection module that incorporates additional contextual information to improve the identification of blurry regions. Considering that modern smartphones are equipped with cameras capable of providing short-exposure images, we develop a blur-aware guided image restoration method that utilizes sharp structural details from short-exposure images, facilitating accurate reconstruction of heavily blurred regions. Furthermore, to restore images realistically and visually-pleasant, we develop a short-exposure guided diffusion model that explores useful features from short-exposure images and blurred regions to better constrain the diffusion process. Finally, we formulate the above components into a simple yet effective network, named ExpRDiff. Experimental results show that ExpRDiff performs favorably against state-of-the-art methods.

Figures (12)

• Figure 1: Deblurring images with real-world local motion blur. The left side shows examples of a long-exposure photo (top) and a short-exposure photo (bottom) of the same scene captured by smart phones with vivo X100 pro. Both existing local blur removal methods, e.g., LBAG ReLoBlur, and state-of-the-art image deblurring methods, e.g., NAFNet NAFNet, do remove local blur well as shown in (a) and (c). In addition, simply using the short exposure image as reference does not solve this problem well, as shown in (b). In contrast to existing methods, we develop an effective ExpRDiff to explore the sharp information from the short-exposure images and blur information from the blurred images for blur removal and use them to guide the diffusion models for better realistic image restoration.
• Figure 2: An overview of the proposed method ExpRDiff. $\mathcal{A}$ is the network to extract features for $R$. $\mathcal{D}$ denotes Blurry-aware guided image restoration, which aims to extract short-exposure image features, CLBDM denotes the context-based local blur detection module. $B$, $R$, $M$, and $H$ are blurry images, short-exposure images, confidence maps, and the result of CLBDM respectively.
• Figure 3: An overview of the proposed context-based local blur detection module to detect the blur regions of the long-exposure blur image. MP denotes the Max-pooling operation.
• Figure 4: Details of the proposed ExpBFusion module.
• Figure 5: Illustration of guided image filtering GIF. (a) Short-exposure reference; (b) blurred input; (c) result of direct guided filtering, which reduces blur but smooths the static background; (d) result of masked guided filtering, preserving background clarity while achieving the same deblurring effect in blurred regions as (c).