ArSL-EdgeBench: Duplicate-Controlled and Edge-Oriented Benchmarking for Static Arabic Sign Alphabet Recognition

Abstract

Reported performance in static Arabic sign alphabet recognition is often difficult to compare because duplicate images, variation across random seeds, robustness to image degradation, and deployment cost are rarely examined in the same experiment. In this study, we introduce ArSL-EdgeBench, a reproducible comparison of four image classifiers on the public ArSL2018 dataset. The initial data set comprised 54,049 images belonging to 32 classes. No partition was generated until duplication detection had been performed based on SHA-256, repeated visual hash inside the same class, and label conflict identification via perceptual hashing. The resulting audit retained 48,093 images that were split into equal training, validation, and test fixed stratified sets according to 70:15:15 proportions. Three models – a lightweight separable CNN MobileNetV3-Small, EfficientNetB0, and ResNet50 were trained using three random seeds each time. EfficientNetB0 came first and was chosen solely based on mean validation macro-F1 score. It demonstrated 94.21% +/- 0.21% mean test accuracy, 94.20% +/- 0.19% macro-F1, and 98.84% top-3 accuracy. Seed-42 variant achieved 94.44% accuracy on the clean test set. Its performance remained almost intact under small angle rotation and reduced illumination, while Gaussian noise decreased macro-F1 from 94.42% to 74.46%. Conversion of the model to Float16 TensorFlow Lite allowed reduction of its size to 8.03 MB with 93.14% accuracy retained; the median single-thread CPU latency was 4.12 ms. Overall, these results demonstrate that EfficientNetB0 is an effective accuracy-efficiency baseline for static alphabet prototypes.  However, they do not create continuous translation or generalize to novel signers since signer identity information is not available in the public dataset.