Amharic Static Alphabet Sign Language Recognition Based on Hybrid Feature Extraction Approach

Sign language is a communication mechanism for hearing-impaired people. Unless hearing individuals learn sign language, communication with hearing-impaired people becomes difficult. To address this communication gap, a sign language recognition system is required. Ethiopian Amharic Alphabet Sign Language (AMASL) differs from that of other countries because the Amharic language is spoken in Ethiopia, and only a few studies have been conducted on AMASL. To bridge these communication gaps, we propose a hybrid feature vector extraction method to develop a suitable recognition model for AMASL. We recorded data from different signers using two different mobile phones in two different cities. Convolutional Neural Networks (CNN) and Histogram of Oriented Gradients (HOG) were fused for feature extraction because CNNs exhibit strong robustness to translation, local deformation, and limited rotational variation due to their convolutional and pooling operations (Lecun et al., 1998). HOG features are effective under low-resolution conditions, strong illumination variation, and small training datasets because they rely on normalized gradient orientation distributions rather than raw pixel values (Dalal and Triggs, 2005). Min–max feature normalization is used to make the features extracted by HOG and CNN comparable. The normalized hybrid feature vector achieves 94.8320 ± 0.9448 mean accuracy ± std using 5-Fold Group cross-validation support vector (SVM).