Advanced Multi-Task Deep Framework for Joint Face Recognition, Age Estimation, Gender Classification, and FineGrained Ethnicity Profiling in Heterogeneous Environments

Abstract

Their use has become commonplace in the field of public safety, safe identity verification, and smart human-computer interaction. High-dimensional changes in wild images like sharp illumination changes, cross domain sensor noise, expressions not cooperating, and difference in demographics, on the other hand, are difficult in modern production systems. This paper fully solves these basic problems by creating an all-encompassing MTL deep convolutional design on top of a better attention-driven backbone network. The system dynamically shares low-level models to carry out identity recognition and simultaneously splits into multiple specialized feature heads to solve the problem of multi-class fine-grained ethnic background and to distinguish into two types of gender identities. To overcome the well known historical algorithmic bias and imbalance in standard training corpora, we employ synthetic demographic equalization, balanced adversarial regularizations and hard focal loss balancing routines. Importantly, we present a new multi-stage inference approach that works better with edge endpoints that don't have a lot of resources. This eliminates the memory and speed issues that sometimes occur with single-layer cascade models. Our proposed model does better than the baseline standards, with a face verification accuracy of 99.42%, an average absolute error (MAE) of 2.14 years in estimating age, a gender detection accuracy of 98.65%, and a classification score of 97.40% across different ethnic groups. The proposed framework is found to be the best in terms of numbers, fairness and hardware reliability from the performance tracking metrics including full confusion matrices and detailed Receiver Operating Characteristic (ROC) curve metrics.