Enhancing missing persons identification through deep learning and GAN based face recognition

This is the authors' abstract. We don't add key points for this paper.

The Lost and Found system is a missing person reporting application that combines a mobile/web reporting platform with an inference only face verification pipeline and a proposed age transformation extension, together with location aware community alerting. The verification pipeline uses MTCNN for face detection and alignment and a pretrained InceptionResNetV1 model (VGGFace2 weights) to extract 512 dimensional, L2 normalized embeddings, compared using cosine similarity and Euclidean distance. A complete, reproducible implementation of this pipeline, including identity disjoint pair generation, five fold cross validation with calibration only threshold selection, fold level Accuracy, Precision, Recall, F1, FAR, FRR, cross fold mean, SD, and 95% CI, a matching rule ablation, a threshold sensitivity check, a representation comparison against a PCA baseline, and CPU latency instrumentation, was implemented and executed end to end on the Olivetti Faces dataset as a development validation experiment, reaching a mean five fold accuracy of 98.64% (SD 1.13 percentage points). Olivetti Faces is a small, controlled, laboratory condition dataset used strictly to confirm that the pipeline, evaluation framework, and statistical procedures operate correctly; it is not one of this study's target missing person relevant benchmarks, and this result is not presented as evidence of performance on unconstrained, crowd sourced imagery. A StyleGAN based age transformation module is proposed as an architectural extension to help bridge large elapsed time gaps in long duration missing person cases; it produces qualitatively plausible age progressed and age regressed images, but its effect on recognition accuracy has not been quantitatively evaluated, and no verified checkpoint or implementation is currently available for a controlled GAN vs no GAN comparison. External validation of the reproducible pipeline on unconstrained, cross dataset target benchmarks, together with the GAN vs no GAN comparison, subgroup fairness analysis, and paired significance testing, remains required and is scoped as immediate future work.