Image restoration method based on GAN and multi-scale feature fusion

• Main Authors: Jin, Xin, Hu, Ying, Zhang, Chu-Yue
• Format: Conference Proceeding
• Language: English
• Subjects: Decoding; Feature extraction; Feature fusion; Fully convolutional neural network; Gallium nitride; GAN; Generative adversarial networks; Generators; Image restoration; Training; VGG-16; WGAN
• ISSN: 1948-9447
• DOI: 10.1109/CCDC49329.2020.9164498

In the research of existing image restore algorithms, defective images and target images exist in pairs, which leads to poor applicability of the algorithm. Aiming at the shortcomings of the existing algorithms, such as low accuracy and poor visual consistency, an image restoration method based on GAN and multi-scale feature fusion was proposed. The generator of this algorithm uses a standard encoder-decoder structure. The encoder based on the VGG-16 full convolutional neural network is used to extract the features of the defect image. Low-dimensional features and high Dimensional features are fused to enrich the input of deep convolutions. Aiming at the problems of large error oscillation amplitude, gradient disappearance or gradient explosion in GAN training, the idea of WGAN is used, and the EM distance is used to simulate the sample data distribution. The similarity between the output restored image and the target image is enhanced by introducing L1 loss. This paper performs experiments on the same face data set. The experimental results show that the algorithm improves the accuracy of the restored image and can generate more realistic restored images.