Robust Sign Language Recognition Using Hybrid Percentile-Based Outlier Aware Dynamic Range Quantization Method

Sign language recognition is an emerging field of computer vision and pattern recognition that enables deaf individuals to communicate with hearing people. With deep learning and advancements in computer vision, computers can automatically recognize hand shapes and motion dynamics, facial expressions, and body postures in sign language. Even though deep neural networks (DNNs) perform well on vision tasks, they consume excessive computational resources and have a large memory footprint. Thus, DNNs cannot be deployed on edge devices. Quantization is one of the most useful model compression techniques for overcoming the limitations of deploying these large models on devices with limited computing resources. This paper proposes an improved post-training quantization method, called the hybrid percentile-based outlier aware dynamic range quantization approach, which sets clipping thresholds layerwise to minimize quantization error. This hybrid quantization approach maintains FP32 precision for input weights, the final output classification layer, and activations, while quantizing only the weights in intermediate layers to achieve computational efficiency. Itworks well in layers with heavy-tailed weight distribution. Experimentation is conducted using benchmark deep learning models, including VGG16, ResNet50, EfficientNetB0, and SqueezeNet. The experiments use a widely used standard ASL Alphabet dataset. The proposed method is robust to outliers and achieves up to 10−11× reductions in model size and 11−122× inference speedups, with less than 1% accuracy drop relative to baseline models with 32-bit floating-point precisions. The findings show that the proposedhybrid quantization method accelerates the inference speed of deep learning models while maintaining optimal performance.