Noise face image super-resolution reconstruction based on multi-information fusion

Abstract

Existing face super-resolution methods are mainly based on an assumption that the input images are noise-free. Their performance degrades drastically when applied to real-world scenarios where the input image is always contaminated by noise. To solve this problem, a multi-information fusion network was proposed, which can reconstruct high-resolution face images with rich details by using two prior information of face parsing image and face attributes. In this method, a coarse SR network was constructed to recover a coarse high-resolution image. Then, the prior information was modeled by an analytic graph estimation network and an attribute analysis and reconstruction network. The reconstruction of high-resolution face images was constrained in pixel space and semantic space. Experimental results show that the proposed method can effectively improve the robustness of the model to noise. On the CelebA data set, the reconstructed image of the proposed method not only have better subjective quality, but also have a peak signal to noise ratio (PSNR) improvement of about 0.2dB compared with the mainstream methods.

Keywords

• noise face image
• super-resolution
• multi-information fusion
• prior guidance
• parsing map
• face attributes