Résumé: Aiming at the problem of image motion blur caused by handheld camera jitter and object motion in the process of collecting photos, a generative adversarial network (GAN) based on feature fusion of back projection is proposed for blind image deblurring. Firstly, the generator network is established by using U-Net structure, and a feature fusion residual block based on back projection is designed according to the error feedback principle, which solves the problem of saving spatial information in U-Net structure. Secondly, the self-attention module is introduced into the generator network to extract the feature map that pays more attention to detail. Finally, the combination of perceptual loss, mean square error loss and relative generative adversarial loss effectively alleviates the mode collapse problem of traditional GAN and improves the stability of model training. The experimental results show that the peak signal-to-noise ratio and structural similarity of this method on GoPro data set are 30.183 dB and 0.941, respectively, and 26.962 and 0.837 on the Kohler dataset, with the shortest running time, which are better than the existing state-of-the-art methods. The restored image is clearer good visual results and richer in texture details, which can effectively improve the image deblurring effect.

Langue: Anglais