Core Concepts
An efficient image deblurring network that leverages selective structured state spaces model to aggregate enriched and accurate local and global features.
Abstract
The paper proposes an efficient image deblurring network called ALGNet that leverages selective structured state spaces model to aggregate enriched and accurate local and global features.
The key components of the proposed method are:
Global Block: Utilizes selective structured state spaces model to capture long-range dependency information efficiently with linear complexity.
Local Block: Models local connectivity using simplified channel attention to address the issues of local pixel forgetting and channel redundancy in the state space equation.
Features Aggregation (FA): Emphasizes the importance of the local block in restoration by recalibrating the weights through a learnable factor when aggregating the global and local features.
The authors design an aggregate local and global block (ALGBlock) that consists of the global and local blocks. Experimental results demonstrate that the proposed ALGNet outperforms state-of-the-art approaches on widely used benchmarks for both image motion deblurring and single-image defocus deblurring, while achieving superior computational efficiency.
Stats
The paper reports the following key metrics:
On the GoPro dataset for image motion deblurring, the proposed ALGNet-B achieves a PSNR of 34.05 dB, outperforming the previous best method NAFNet-64 by 0.43 dB.
On the HIDE dataset for image motion deblurring, the proposed ALGNet achieves a PSNR of 31.68 dB, outperforming the previous best method Restormer-local by 0.19 dB.
On the RealBlur-R dataset for real-world deblurring, the proposed ALGNet achieves a PSNR of 41.21 dB, outperforming the previous best method MRLRFNet by 0.29 dB.
On the DPDD dataset for single-image defocus deblurring, the proposed ALGNet achieves a PSNR of 26.45 dB, outperforming the previous best method IRNeXt by 0.15 dB.
Quotes
"Our ALGNet-B shows 0.43 dB performance improvement over NAFNet-64 [4] on GoPro [20]."
"Even though our network is trained solely on the GoPro [20] dataset, it still achieves a substantial gain of 0.19 dB PSNR over Restormer-Local [16] on the HIDE [43] dataset."