Deep learning with attention mechanism for iterative face super-resolution

Abstract

Face images are widely used in many applications, such as face recognition and face identification. Regarding security, face identification is used to track the crimes. However, the camera's low resolution and environmental degradation problem hinders the face application's performance. In this thesis, we study face image super-resolution to restore the image from low-resolution to high-resolution.

We proposed deep learning with an attention mechanism for iterative face super-resolution that included an image super-resolution network and face alignment network combined. The input low-resolution image is enlarged into a super-resolution face image. Then, the image has repeatedly estimated the alignment to enhance the super-resolution image.

The experiment was conducted on well-known facial datasets. Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM) are measured for objective performance evaluation. The performance of the proposed method is compared with bicubic interpolation and other referenced methods. The experimental results demonstrate that the proposed method has the highest performance compared with other reference methods.