| 摘 要: 分裂算法在图像去模糊领域应用广泛,但现有基于CNN架构的算法,难以处理具有复杂纹理结构的先验。因此,引入了基于通道注意力的Transformer架构作为深度先验降噪器,并将其融入半二次分裂算法(HQS)。具体来说,通过HQS算法,将图像去模糊的最大后验概率估计问题解耦为两个单独的子问题。其中,数据项子问题利用快速傅里叶变换(FFT)求解,先验项子问题通过基于通道注意力的Transformer架构求解。实验结果表明,在BSD300测试集上,该算法与FDN、DWDN、IRCNN、DPIR4种方法相比,PSNR和LPIPS指标更具优势,证明了其有效性。 |
| 关键词: 深度学习 图像去模糊 半二次分裂算法 Transformer |
|
中图分类号: TP391
文献标识码: A
|
|
| Image Deblurring Method Based on Channel Attention and Half-Quadratic Splitting Algorithm |
|
YU Kai1, WANG Meiliang1,2, SHEN Yang1,2
|
(1. School of Information Science and Engineering, Zhejiang Sc-i Tech University, Hangzhou 310018, China;
2.School of Mathematics and Computer Science, Lishui University, Lishui 323000, China)
2574406777@qq.com; 18767883365@163.com; tlsheny@126.com
|
| Abstract: Splitting algorithms have been widely applied in image deblurring, yet existing CNN-based architectures struggle to handle priors with complex texture structures. To address this, we introduce a Transformer architecture with channe-l wise attention as a deep prior denoiser and integrate it into the Hal-f Quadratic Splitting (HQS) algorithm.Specifically, the HQS algorithm decouples the maximum a posteriori probability estimation problem of image deblurring into two distinct subproblems. The data fidelity subproblem is solved via Fast Fourier Transform (FFT),while the prior subproblem is addressed by the channel attention-based Transformer architecture. Experimental results demonstrate the superiority of our method over four existing approaches (FDN, DWDN, IRCNN and DPIR), as evidenced by higher PSNR and LPIPS metrics on the BSD300 test set. |
| Keywords: deep learning image deblurring hal-f quadratic splitting |
