A Multi-Scale Gabor Filter Bank Approach for Handshape Feature Extraction in Vision-based Sign Language Recognition Systems

Abstract

Sign language is the primary mode of communication for millions of Deaf and hard-of-hearing individuals, yet the absence of affordable, accurate, and real-time translation tools continues to isolate this community from mainstream communication channels. This paper reviews the principal issues and challenges confronting vision-based Sign Language Recognition (SLR) systems - spanning data scarcity, linguistic complexity, occlusion, and real-time deployment constraints - and proposes a multi-scale Gabor filter bank for handshape feature extraction as a lightweight alternative to purely deep-learning pipelines. Filter wavelengths (λ) are systematically mapped to kernel sizes using the standard Gabor relation σ = 0.56λ, producing three operating bands - fine (15×15 to 27×27), transition (31×31 to 37×37), and coarse (41×41 to 69×69) - each tuned to a distinct anatomical feature scale, from fingernail edges to full hand and body silhouette. We discuss the design rationale, computational trade-offs, and integration of this multi-scale representation within a broader recognition pipeline, and outline directions for experimental validation.