SignNet-nano: efficient sign language recognition for real-time edge deployment

Abstract

Real-time american sign language (ASL) alphabet recognition can enable inclusive communication on next-generation edge devices like virtual reality (VR) headsets and smart glasses. However, existing models are often too computationally demanding for such platforms, resulting in fast battery depletion and degraded performance. To address this challenge, we introduce SignNet-Nano, an ultra-lightweight convolutional neural network (CNN), tailored specifically for efficient ASL alphabet recognition on constrained edge hardware. The proposed model integrates depthwise separable (DS) convolutions, squeeze-and-excitation (SE) attention, and global average pooling (GAP) to achieve high accuracy with a compact architecture comprising only < 20K parameters and < 0.1MB model size. Unlike prior work, SignNet-Nano is explicitly designed with deployment efficiency as a primary goal, enabling real-time inference with minimal performance degradation while requiring significantly less energy, latency, computation, and memory usage than state-of-the-art (SOTA) lightweight edge models. We profile inference performance in terms of Floating Point Operations (FLOPs), inference latency, frames per second (FPS), and memory footprint on a diverse set of platforms, including three edge devices - Jetson Nano, Jetson Xavier NX, and Raspberry Pi 4 - as well as three high-performance systems - Apple Mac M3 (MPS backend), NVIDIA RTX 4070 GPU, and an Intel Core i7 12th Gen CPU. Experimental results show that SignNet-Nano achieves classification accuracy within 1% of the best-performing baseline while reducing inference time (i.e., latency) and FLOPs by up to 73.7% and 97.1%, respectively, while improving the energy efficiency by up to 79.3%.