Date of Award

2026-08-01

Degree Name

Master of Science

Department

Computer Science

Advisor(s)

Olac Fuentes

Abstract

Facial age estimation supports law enforcement via image-based, age-filtered queries, age-progressive re-identification, and bulk record labeling, where prediction accuracy determines if the resulting decisions can be trusted. State-of-the-art models excel on web imagery but incur higher error on mugshots due to domain shift between the professionally lit, filtered, and posed web photographs used during pre-training and the uniform backgrounds, uncooperative expressions, and decades of evolving capture technology found in mugshot collections. We address this gap by adapting SwinFace - a state-of-the-art multi-task Swin Transformer with public code and pretrained weights, trained on color face imagery for face recognition, facial expression recognition, age estimation, and attribute estimation - to the mugshot domain. We evaluate this adaptation on two independent mugshot datasets to assess cross-modality generalization: the grayscale NIST-18 benchmark (1,574 frontal images) and NE-Mugshot20K, a proprietary color dataset of 20,643 frontal images spanning 1996-2025. The adaptation comprises: (A) image pre-processing via face detection, alignment, and contrast enhancement; (B) layer depth unfreezing; (C) learning rate and scheduler adjustment; (D) joint age-gender fine-tuning; (E) test-time augmentation (TTA); and (F) post-training linear calibration. For NIST-18's grayscale imagery, a learned 1 x 1 convolutional expansion replaces static channel replication. Ablation studies attribute the largest gains to pre-processing and fine-tuning depth; the combined methodology minimizes age estimation error to 3.11 ± 0.03 years on NIST-18 (67.87% reduction from zero-shot) and 3.18 ± 0.07 years on NE-Mugshot20K (84.00% reduction). Joint gender-age training, explored as a regularizer, increased error (+0.03/+0.01 years, NIST-18/NE-Mugshot20K) while restoring gender accuracy to 98.77% and 98.95%. Global post-training calibration reduced bias without MAE gain, while per-age-group calibration yielded no net benefit, consistent with a non-linear age-dependent bias pattern. These results show that structured adaptation achieves competitive performance across distinct forensic imaging modalities.

Language

en

Provenance

Received from ProQuest

File Size

76 p.

File Format

application/pdf

Rights Holder

Jorge Alejandro Pacheco Roque

Download

Share

COinS