•  
  •  
 

Abstract

This research proposes and validates a novel Hybrid DenseNet121-InceptionV3 architecture for robust multi-output demographic classification from facial images, with a focus on optimizing both time efficiency and parameter utilization. The model was trained on a curated subset of the UTKFace dataset comprising 4,000 images, balanced across four racial groups with equal gender distribution. Our hybrid approach processes images through parallel feature extractors, strategically combining DenseNet121's parameter-efficient feature reuse with InceptionV3's multi-scale analysis capabilities. This architectural synergy achieves superior performance while maintaining computational efficiency. Through rigorous 3-fold cross-validation, the hybrid model demonstrated significant advantages in accuracy and training time, achieving 98.35% race classification accuracy, 97.60% gender accuracy, and 96.22% combined accuracy—outperforming all standalone baseline models by 2–3%. The model's parameter-optimization strategy enabled faster convergence during training while reducing computational overhead, making it suitable for real-time applications. Notably, the architecture attained this enhanced performance without increasing parameter complexity, demonstrating that intelligent model fusion can simultaneously optimize both temporal efficiency and parameter utilization. These findings establish a new benchmark for developing computationally efficient deep-learning models that balance high accuracy with practical deployment requirements in multi-output classification tasks.

Download

Share

COinS