Comparative Evaluation Of FP32 And INT8 Quantization For Edge Device Real-Time Indian Sign Language Recognition

The high memory and computational requirements of standard floating-point architectures make it difficult to deploy Deep Learning models on edge devices with limited resources. The Post-Training Quantization (PTQ) methods used with a MobileNetV2 architecture for real-time Indian Sign Language (ISL) recognition are compared in this paper. We show that the baseline 32-bit floating-point (FP32) model can be converted to an 8-bit integer (INT8) format, which significantly reduces the model footprint and improves inference latency without significantly degrading accuracy. The INT8 quantized model achieves a 72.4% reduction in memory size (from 9.52 MB to 2.63 MB) and a 31.7% increase in inference speed (from 14.40 ms to 9.83 ms per frame), according to experimental results on a 27-class ISL dataset. Importantly, compared to the FP32 baseline, the quantized model maintains a strong validation accuracy of 98.60%, with an accuracy decline of less than 1%. These results confirm that INT8 quantization is effective in enabling offline, high-framerate computer vision applications on low-power edge hardware.