Image blur processing technology based on multi-scale feature network and cross-stage attention mechanism
Yanjun Ma · IET conference proceedings. · 2023
Aiming at the problem that traditional image deblurring techniques are not effective in dealing with high blurring images, a novel image deblurring method based on multi-scale feature network and cross-stage attention mechanism is proposed. By combining multi-scale feature extraction and cross-stage attention mechanism, the method aims to enhance the acquisition of spatial characteristic information of images while reducing feature loss to improve the effect of image deblurring. In the experiments on the GoPro dataset, compared with the traditional single-scale method, the investigated two-scale processing method improves the peak signal-to-noise ratio by 0.843 dB and grows the structural similarity by 0.0116, while the three-scale processing method further increases the PSNR and SSIM by 0.756 dB and 0.0093, respectively, in comparison with the two-scales.On the Kohler dataset, the present method improves 0.04dB in PSNR and 0.013dB in SSIM compared to other mainstream generative adversarial network algorithms.Compared to the scaled recurrent network, the research method outperforms the Kohler dataset by 0.31dB in PSNR and 0.018dB in SSIM performance.The results show that the combination of multiscale feature network and cross-stage attention mechanism has significant advantages in improving the recovery effect of blurred images and provides an efficient new approach in the field of image processing, especially showing strong potential in dealing with highly blurred images.