Research on gesture recognition algorithm based on YOLOX
Dongyu Liu, Di Xiao, Yue Lu, Shan Wu · 2025
Sign language aids the hearing - impaired. To solve their communication problems, this article presents an enhanced YOLOX model for sign - language recognition. Targeting YOLOX's high parameter count and computational complexity, an SG_YOLOX architecture is proposed. It integrates lightweight ShufflenetV2 and Ghost PAN, effectively solving these issues. ShufflenetV2 extracts features, feeding three - scale feature maps into Ghost PAN, which combines high - level semantic and low - level localization info, aggregating parameters from different detection layers.With a 50% probability, Mosaic and Mixup data augmentation are used. Leaky ReLU is the activation function, Focal Loss for the classification branch, and DIOU Loss for the regression branch. The Head outputs the target's location and category.Validated on the American Sign Language alphabet dataset, compared to the original YOLOX, the proposed model reduces parameter count by 2.6 times, computational complexity by 4.5 times, and increases mAP by 4.9 percentage points to 96.4%. This model effectively cuts parameters and complexity, boosts sign - language recognition accuracy, and shows good robustness.