.png){kind=icon}
.png){kind=icon}
Return
SMem-Diff: A Simple Memory-Augmented Diffusion Model for Effective Video Deblurring on Cloud-Edge Servers
.png){kind=icon}
PRE
AI
DOI：10.1109/tcsvt.2026.3691297.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deep learning–based video deblurring can exploit inter-frame information to remove blur artifacts in videos captured by cameras on edge devices, thereby enhancing video quality. Most existing methods adopt sliding window mechanism, recurrent feature propagation, transformers, or their combinations to leverage inter-frame information in video sequences. However, these methods typically process only a subset of adjacent blurry frames at a time and cannot effectively extract useful information from distant frames, thus failing to fully leverage the available information from the entire input video. Moreover, most of them are based on regression models, where the network fundamentally relies on highly degraded blurry input frames, causing the quality of the output frames to degrade as the input frames deteriorate. To address these issues and achieve finer-grained deblurring, we construct a simple yet effective memory bank for edge servers to filter and store temporal frame information, and design a memory update algorithm based on usage frequency to further exploit the advantages of memory networks. Additionally, we introduce the diffusion priors by combining a diffusion network with the designed hierarchical dense attention module to further alleviate the model’s high sensitivity to degraded input frames. Qualitative and quantitative experiments show that our proposed SMem-Diff exhibits superior performance on the GOPRO, DVD, and BSD datasets, with quantitative comparison result on the GOPRO dataset exceeding STCT+ by over 0.4dB in PSNR.
Keywords:
Video deblurring
temporal information
memory bank
diffusion network
edge servers
Journal
IF:
11.1
Papers:
612
Citations:
3.1W
View more organizations