Design and Evaluation of a Static Filipino Sign Language Alphabet Recognition System Using Support Vector Machine

Filipino Sign Language (FSL) is the recognized form of communication for the deaf and hard-of-hearing individuals under Republic Act No. 11106. Despite this, Filipino Sign Language remains underrepresented in technological research and development. This study aims to develop and evaluate an AI-based system for recognizing static FSL alphabet hand gestures. The study follows a quantitative and experimental research design. The study implements hyperparameter optimization of the regularization parameter of the Support Vector Machine (SVM) utilizing Histogram of Oriented Gradients (HOG) feature extraction. The regularization parameter controls the trade-off between the margin and minimizing classification errors. The system is trained and evaluated using a public dataset available in Kaggle consisting of 11,700 preaugmented hand gesture images representing the 26 letters of FSL alphabet. The preprocessing techniques used are grayscale conversion, otsu’s thresholding technique, and morphological operations to enhance the hand segmentation, afterwards HOG features are extracted and used as an input to the SVM classifier. Two data split configurations (50:50 and 80:20) employed to assess the model generalization and robustness. To compare the performance of the SVM model, a Convolutional Neural Network (CNN) is implemented as a baseline for performance comparison. Results show that the SVM model achieved a maximum accuracy of 98.55% using a 80-20 training validation split outperforming its performance under a 50:50 configuration, which yields 91.86% accuracy. The baseline CNN model achieves a comparable accuracy of 97.99%, indicating that non-neural network techniques can perform as effectively as deep learning models for static gesture tasks. Moreover, confusion matrix analysis reveals that misclassifications primarily occur among visually similar gestures. Overall, SVM classifications for static FSL recognition are highly effective under 80:20 test validation split and C = 0.5 reaching an accuracy of 98.55%. The study demonstrates the potential of computationally efficient models for use in accessible learning tools, while at the same time providing a baseline for further work in dynamic and multimodal sign language recognition systems.