Transfer Learning-Based Distance-Adaptive Global Soft Biometrics Prediction in Surveillance
Sonjoy Ranjon Das, Henry Onilude, Bilal Hassan, Preeti Patel, Karim Ouazzane
Electronics (2025), Volume 14, Issue 18, 3719
Abstract
Soft biometric prediction in surveillance systems faces significant challenges when attempting to maintain accuracy across varying subject distances. This research presents a novel transfer learning-based distance-adaptive approach for predicting global soft biometrics including age, gender, and ethnicity in surveillance environments. Using EfficientNetB3 with distance-adaptive multi-task deep learning, our method achieves robust performance across distances ranging from 4 meters to 10 meters.
Key Highlights
- ▸95% gender classification accuracy at 4 meters
- ▸85% gender accuracy maintained at 10 meters
- ▸65%+ ethnicity recognition across all distances
- ▸Age estimation MAE of 1.1-1.5 years
- ▸Tested on 19,236 samples from FVG and MMV datasets
- ▸Distance-adaptive multi-task learning approach
- ▸EfficientNetB3 transfer learning architecture
Keywords
Citation
Das, S.R.; Onilude, H.; Hassan, B.; Patel, P.; Ouazzane, K. Transfer Learning-Based Distance-Adaptive Global Soft Biometrics Prediction in Surveillance. Electronics 2025, 14, 3719. https://doi.org/10.3390/electronics14183719