ANNLR-FSR: Artificial Neural Network-Based Locality Regularization for Face Super-Resolution

In electronic surveillance, it is quite challenging to improve the resolution of consumer gadget's captured facial images which are often degraded by irregular noise. Although, there has been lot of research in the field of face super-resolution (FSR) published over past 20 years, but only a small percentage of them have concentrated on the issue of noisy FSR. Moreover, they reconstruct the noisy low-resolution (LR) faces without bothering about the actual noise level, which is essential for achieving state-of-the-art performance. Therefore, in this paper, an artificial neural network (ANN) based locality regularization technique for FSR is proposed named as ANNLR-FSR. In this, a noise level estimation model based on ANN is presented that precisely predicts the actual level of noise present in the test face. Further, the estimated noise is associated with the LR training images to make them structurally similar to the test face. This helps in obtaining noise-resistant reconstruction coefficients and achieving good super-resolution performance. Experiments are conducted on FEI, CelebA face datasets, and locally captured real-life faces to verify the performance of the proposed ANNLR-FSR technique. The obtained results evidently indicate that the proposed technique outperforms other compared state-of-the-art FSR methods.