Blind Image Deblurring Based on Adaptive Redescending Potential Function and Local Patch Fidelity Term
Lulu Zhang,Qiyu Jin,Guoliang Zhao,Caiying Wu
DOI: https://doi.org/10.1007/s11760-024-03512-7
IF: 1.583
2024-01-01
Signal Image and Video Processing
Abstract:Blind image deblurring is a fundamental and important task in the field of computer vision. With the continuous progress of technology, blind image deblurring methods have achieved remarkable development and wide application. In early research, methods based on the L_0+X paradigm have successfully solved the blind image deblurring problem to some extent. Recently, a new method based on the redescending potential function has also come into prominence. This new method is formally more concise, but the general redescending potential function does not adapt to different regional features of the image. To address this issue, we introduce an adaptive redescending potential function. This function adapts to different structural features based on the magnitude of the total curvature in different regions of the image. Additionally, we introduce a local block fidelity term to consider the difference in gradient information, which is not involved in previous blind deblurring methods. The efficacy of the proposed method is demonstrated through experimental results on benchmark datasets and real blurred images. Our method achieved improvements in PSNR results by 0.28 dB and 0.79 dB on the Köhler and Levin datasets, respectively. Similarly, on the Lai dataset, our method enhanced the PSNR by 0.77 dB and the SSIM by 0.0374. Moreover, the self-supervised model induced by our proposed method achieved a PSNR result on the Levin dataset that is 0.47 dB higher than other advanced self-supervised models. These results indicate that the method is effective and produces notable results.