Predicting the Future Appearances of Lost Children for Information Forensics with Adaptive Discriminator-Based FLM GAN

Abstract

This article proposes an adaptive discriminator-based GAN (generative adversarial network) model architecture with different scaling and augmentation policies to investigate and identify the cases of lost children even after several years (as human facial morphology changes after specific years). Uniform probability distribution with combined random and auto augmentation techniques to generate the future appearance of lost children’s faces are analyzed. X-flip and rotation are applied periodically during the pixel blitting to improve pixel-level accuracy. With an anisotropic scaling, the images were generated by the generator. Bilinear interpolation was carried out during up-sampling by setting the padding reflection during geometric transformation. The four nearest data points used to estimate such interpolation at a new point during Bilinear interpolation. The color transformation applied with the Luma flip on the rotation matrices spread log-normally for saturation. The luma-flip components use brightness and color information of each pixel as chrominance. The various scaling and modifications, combined with the StyleGan ADA architecture, were implemented using NVIDIA V100 GPU. The FLM method yields a BRISQUE score of between 10 and 30. The article uses MSE, RMSE, PSNR, and SSMIM parameters to compare with the state-of-the-art models. Using the Universal Quality Index (UQI), FLM model-generated output maintains a high quality. The proposed model obtains ERGAS (12 k–23 k), SCC (0.001–0.005), RASE (1 k–4 k), SAM (0.2–0.5), and VIFP (0.02–0.09) overall scores.