关键词: 生成对抗网络, 高频分量, 超分辨率, 判别器, 损失函数

Abstract:

In order to solve the problem of low peak signal-to-noise ratio and structural similarity of reconstructed face images， a super-resolution reconstruction method based on generative adversary-network is proposed. A two-branch generator network is created， the shallow features of the image are obtained by using the feature extraction module in the network， the high-frequency branches of the image are reconstructed， the high-frequency data of the low-resolution image is trained by the deep residual network， and the face image data is mapped. Stop updating fixed generator parameters， use relative discriminator network to identify the authenticity of face image reconstruction， introduce attention mechanism to optimize the reconstruction performance of generative adversarial network， evaluate the reconstruction reliability by perception loss function， and realize the super-resolution reconstruction of face. The experimental results show that the proposed method can effectively improve the ability of facial detail reconstruction， and the subjective and objective evaluation indexes are excellent. The subjective evaluation results show that the reconstructed facial images with the proposed method have clearer detailed data， more three-dimensional facial contours， less distortion rate， and more natural and real external features. The objective evaluation results show that the peak signal-to-noise ratio of the proposed method reaches 44.6dB and the structural similarity reaches 0.74， which fully demonstrates the robustness of the proposed method and provides an effective reference for optimizing the application performance of face recognition.

Key words: generate adversarial networks, high frequency component, super resolution, discriminator, loss function