Spatially-Attentive Patch-Hierarchical Network with Adaptive Sampling for Motion Deblurring

TL;DR

This work tackles motion deblurring in dynamic scenes by introducing a pixel-adaptive, content-aware architecture that jointly leverages global attention and local, pixel-dependent filtering in a three-stage, patch-based hierarchy. A non-uniform, learnable sampling strategy concentrates computation on hard-to-restore regions, guided by reinforcement learning to select informative pixels, and GT supervision is applied progressively to refine restoration. The method integrates cross- and self-attention with a pixel-dependent filtering module to realize a robust global-local fusion, achieving state-of-the-art or competitive results on GoPro, HIDE, and RealBlur benchmarks while maintaining efficiency. Overall, the approach provides interpretable, adaptive restoration where blur varies across spatial locations, with potential applicability to broader image restoration tasks.

Abstract

This paper tackles the problem of motion deblurring of dynamic scenes. Although end-to-end fully convolutional designs have recently advanced the state-of-the-art in non-uniform motion deblurring, their performance-complexity trade-off is still sub-optimal. Most existing approaches achieve a large receptive field by increasing the number of generic convolution layers and kernel size. In this work, we propose a pixel adaptive and feature attentive design for handling large blur variations across different spatial locations and process each test image adaptively. We design a content-aware global-local filtering module that significantly improves performance by considering not only global dependencies but also by dynamically exploiting neighboring pixel information. We further introduce a pixel-adaptive non-uniform sampling strategy that implicitly discovers the difficult-to-restore regions present in the image and, in turn, performs fine-grained refinement in a progressive manner. Extensive qualitative and quantitative comparisons with prior art on deblurring benchmarks demonstrate that our approach performs favorably against the state-of-the-art deblurring algorithms.

Figures (10)

• Figure 1: Example of non-uniform sampling. Sampled pixels are shown in red. Best viewed when zoomed-in.
• Figure 2: Overall architecture of our proposed network. CA block represents cross attention between different levels of encoder-decoder and different levels. All the resblock contains one content aware processing module. '+' denotes elementwise addition. We include skip connection between the encoder and decoder, which is described in Fig. \ref{['fig:content_aware']} for the non-uniform sampling.
• Figure 3: Illustration of our content-aware processing module. The upper and lower branch show self-attention and PDF module, respectively.
• Figure 4: Visual comparisons of deblurring results on images from the GoPro test set gopro2017.
• Figure 5: Visual comparisons of deblurring results on images from the HIDE test set.