A HYBRID AI-DRIVEN FRAMEWORK FOR HIGH-PERFORMANCE FACIAL SKETCH RECOGNITION IN BIOMETRIC IDENTIFICATION SYSTEMS

Objective: Facial sketch recognition has been an issue of concern in biometrics in forensics as there is low accuracy of sketch-to-photo recognition of sketches in heterogeneous visual conditions. Method: In this paper, it is suggested that a hybrid AI-based framework can incorporate the use of a CNN backbone to extract local faces to model global features and a Transformer encoder to model global features and an adversarial domain adaptation component to align cross-domain embeddings. Transfer learning, data augmentation, hard-negative mining and joint optimization objective (comprising cross-entropy, triplet, contrastive, and adversarial losses) are applied to the framework to train it on CUFS, CUFSF, and IIIT-D sketch datasets. Rank-1 accuracy, precision, recall and ROC-AUC are the evaluations used to evaluate the experiment. Results: Using the proposed framework, in the illustrative benchmark set-up indicated in this manuscript, the framework attains a Rank-1 accuracy of 98.6% on CUFS and 91.8% on CUFSF and 84.7% on IIIT-D and achieves a ROC-AUC of 0.976, which outperforms the corresponding representative prior baselines. Novelty: These findings suggest that hybrid local-global-domain learning can be a powerful and scalable method to apply to real-world forensic sketch recognition, and it has great potential to be implemented into law-enforcement, and biometric identification systems.