A low-complexity method for efficient depth-guided image deblurring

TL;DR

The paper tackles depth-guided image deblurring on resource-constrained devices by modeling blur as a convolution $y = k \circledast x$ and introducing EDIBNet, a lightweight network that operates in the wavelet domain using a 2-level Haar DWT to separate coarse structure into $LL^{(2)}$ and discard less-informative high-frequency content. It fuses mobile LiDAR depth through lightweight adapters within an Efficient U‑Net backbone, enabling structure-aware restoration with minimal overhead. The authors demonstrate that EDIBNet achieves competitive PSNR/SSIM while reducing FLOPs, runtime, and memory by up to two orders of magnitude compared to state-of-the-art models, particularly on edge devices such as the NVIDIA Jetson Orin Nano. Ablation studies validate the choice of a 2-level decomposition, Haar basis, and the proposed depth adapters, and the results show real-time performance (around 0.2 s per high-resolution frame) with robust depth integration, highlighting practical impact for mobile multimodal imaging.

Abstract

Image deblurring is a challenging problem in imaging due to its highly ill-posed nature. Deep learning models have shown great success in tackling this problem but the quest for the best image quality has brought their computational complexity up, making them impractical on anything but powerful servers. Meanwhile, recent works have shown that mobile Lidars can provide complementary information in the form of depth maps that enhance deblurring quality. In this paper, we introduce a novel low-complexity neural network for depth-guided image deblurring. We show that the use of the wavelet transform to separate structural details and reduce spatial redundancy as well as efficient feature conditioning on the depth information are essential ingredients in developing a low-complexity model. Experimental results show competitive image quality against recent state-of-the-art models while reducing complexity by up to two orders of magnitude.

Figures (4)

• Figure 1: Proposed EDIBNet architecture. An efficient encoder-decoder neural network operates in the low-frequency wavelet sub-bands. Efficient adapters are added on each level of the decoder part to modulate image features with depth features.
• Figure 2: Architecture of the proposed efficient Adapter. The module takes as input both image features and depth features. Each input is first normalized and passed through lightweight bias adjustment layers. The features are then concatenated and processed through a chunking and spatial conditioning mechanism, which generates prompt signals that adaptively modulate the image features based on geometric context. This design enables efficient and structure-aware fusion of depth cues into the visual pipeline with minimal computational overhead.
• Figure 3: Deblurring image quality (PSNR) against Runtime (sec.).
• Figure 4: Qualitative results.