Reconstruction of Low-Resolution Facial Images Using Improvements to the FSRCNN Model with Parallel Feature Extraction Layers and Residual Connections

Enhancing the quality of reconstructing low-resolution facial images poses a significant challenge in various applications, particularly in CCTV-based surveillance. This study develops a Fast Super-Resolution Convolutional Neural Network (FSRCNN) model with the addition of parallel feature extraction layers and residual connections from these layers to the expanding layer to improve the reconstruction quality of faces from low resolution. The proposed model is tested using two datasets: CelebA and CCTV footage. The tests are conducted for two reconstruction scales, namely from LR to HR and from LR to intermediate. The results demonstrate that the proposed model consistently outperforms the standard FSRCNN. In the LR to HR reconstruction, the proposed model achieves an average PSNR improvement of 1.24 dB and an SSIM increase of 0.0584 on the CelebA dataset. For the CCTV dataset, an average PSNR enhancement of 0.52 dB and an SSIM improvement of 0.025 are attained. While this model significantly enhances image quality, several limitations were identified regarding performance on more varied CCTV images and the higher complexity of the model.