A Hybrid Feature Fusion Approach for Gender Detection from Palm Image

Abstract

Gender classification from hand images is a useful soft-biometric task in forensic analysis, access control, and demographic analytics, particularly when the face is occluded or unavailable. Existing palm-based methods usually rely either on handcrafted texture descriptors or on a single deep convolutional network, and each family captures only part of the discriminative information in a palm. This paper proposes a hybrid feature fusion framework that combines (i) handcrafted texture and shape descriptors (Histogram of Oriented Gradients, multi-scale Local Binary Patterns, Gabor filter-bank statistics, and geometric palm measurements) with (ii) deep features extracted by a fine-tuned convolutional backbone. The two feature streams are normalized, projected into a shared space, and merged by a gated attention fusion module that learns how much to trust each stream for every input image. A lightweight classifier operates on the fused representation. We describe a subject-independent evaluation protocol on the public 11K Hands dataset, an ablation plan that isolates the contribution of each feature family and fusion strategy, and an explainability analysis based on Grad-CAM. The framework is designed to be reproducible, computationally modest, and robust to illumination and pose variation. Result tables in this manuscript are provided as templates to be completed with the outputs of the accompanying experimental pipeline.