Identity-aware Pseudo-Supervision Face Image Super-resolution for Surveillance Scenarios

Generic super-resolution models adopt a predefined degradation kernel (e.g., Gaussian) or estimate the degradation kernel from low-resolution images to downsample high-resolution images as the training data, because there are few low-resolution and high-resolution image pairs. However, the low-resolution images captured in surveillance scenarios are severely impacted by capture distance, illumination, occlusion, etc., so the degradation kernel is too complicated to predefine or estimate. Meanwhile, identity preservation is significant for face image super-resolution. To this end, we propose identity-aware pseudo-supervision face image super-resolution in real-world scenarios, especially surveillance scenarios. We utilize Generative Adversarial Nets (GAN) to degenerate high-resolution images to follow the distribution of low-resolution images captured by surveillance cameras. The synthesized low-resolution images and corresponding high-resolution compose paired training data with pseudo-supervision. Furthermore, to preserve the identity during the degeneration, a face image quality assessment module is plugged into the degeneration module, which balances the degradation and identity preservation. We conduct extensive experiments on multiple benchmarks to verify the effectiveness of the proposed method.