GLAGAN image inpainting algorithm based on global and local consistency

• Published in: 2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC) Vol. 9; pp. 646 - 650
• Main Authors: Li, Xiaoli, Zhou, Shuailing
• Format: Conference Proceeding
• Language: English
• Published: IEEE 11.12.2020
• Subjects: Convolution; Correlation; cross attention module; deep learning; Image edge detection; image inpainting; image processing; Image restoration; Information technology; Maintenance engineering; Training
• DOI: 10.1109/ITAIC49862.2020.9339126

Image inpainting is an important part in the field of image processing, and its purpose is to complete the damaged area according to the pixel information not lost in the image. At present, although the image inpainting algorithm based on deep learning can repair the missing area, there are still problems that the effective information of the far area cannot be obtained, and the edge of the repair area is blurred or even distorted. In view of the above problems, this paper proposes a GLAGAN image inpainting algorithm based on global and local consistency, so that the repaired image can be semantically consistent globally and locally. In the generation network, the dilated convolution is used to initially repair the missing area, and the cross attention module is used to obtain the correlation between the repaired area and the known area, and the feature weight is calculated to further repair the damaged image. Then through the global discriminator and local discriminator to conduct adversarial training to improve the consistency of the image repair results. Experimental results show that the repair effect of the algorithm is more real and natural, and it has been further improved in both subjective evaluation and objective evaluation.