Arabic Sign Language Alphabet Recognition Using Transfer Learning: Evaluation, Ablation, and Deployment

Arabic Sign Language (ArSL) is one of the most widely used sign languages among the hearing-impaired community in Arabic-speaking regions. Yet, the automated recognition of its alphabet remains a critical challenge for assistive technology development. This study presents a transfer learning classification model for Arabic Sign Language Alphabets (ArSLA) based on the InceptionV3 convolutional neural network architecture. The research contribution is a rigorously evaluated, reproducible recognition pipeline that advances diagnostic depth beyond prior studies through ablation testing, per-class F1-score analysis, and confusion pattern interpretation. The ArSL2018 dataset, comprising 54,049 images distributed across 32 Arabic alphabet classes, was used for training and evaluation. The model integrates transfer learning from ImageNet-pretrained weights, Global Average Pooling, a 256-unit dense layer with ReLU activation, dropout regularization (rate: 0.4), and a 32-class softmax output layer. Training employed the Adam optimizer with adaptive learning rate scheduling and early stopping callbacks. Model evaluation was conducted using a stratified 80:20 train-test split replicated across five independent runs with different random seeds, yielding a mean test accuracy of 97.41% ± 0.31% and a best single-run test accuracy of 97.68%, outperforming all previously reported models on the same benchmark dataset. An ablation study confirmed the independent contributions of transfer learning, data augmentation, and dropout regularization. A real-time prototype was implemented using OpenCV at 3.06 FPS on CPU hardware. These findings establish InceptionV3-based transfer learning as a strong and reproducible baseline for Arabic sign language assistive technology.

@article{Maarouf2026,
  title     = {Arabic Sign Language Alphabet Recognition Using Transfer Learning: Evaluation, Ablation, and Deployment},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {6},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Abdelfatah Maarouf and Otman Maarouf and Abdelaali Benaiss and Rachid El Ayachi and Mohamed Biniz},
  doi       = {10.14569/IJACSA.2026.0170657},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170657}
}