Facial Classification for Pensioner Verification Using Gradient Features and a Feed Forward Neural Network

Reliable pensioner verification requires a clear connection between facial evidence and an enrolled identity. This study examines a gradient-feature and feed-forward neural-network pipeline developed for pension administration in Enugu North, Nigeria. Facial images were acquired, processed by Gaussian smoothing, represented using histograms of oriented gradients, and analyzed within a MATLAB-based workflow. The collection description records 4,030 images from 2018–2021; the available development confusion matrices contain 699 classified observations, partitioned into 489 training, 105 validation and 105 test cases. Reanalysis of these matrices yielded accuracies of 96.93%, 96.19% and 97.14%, respectively. On the test partition, Class 2 precision, recall and F1-score were 94.87%, 97.37% and 96.10%; balanced accuracy was 97.19%. Best validation cross-entropy was 0.041598 at epoch 8, while the recorded run stopped at epoch 14. A separate application table contained 18 True-coded outcomes in 20 cases. The results indicate effective discrimination within the documented two-class experiment, with similar performance across development partitions. The study formalizes the processing model and establishes a count-based evaluation that separates model discrimination from operational identity verification. Wider deployment requires clarification of class semantics, linkage between the collection and development subsets, and independently adjudicated identity claims.