Design and Implementation of Lightweight Vision Transformer for Low-Power Edge Devices

Vision Transformers (ViTs) have emerged as a paradigm shift in computer vision, surpassing convolutional neural networks (CNNs) in tasks requiring global context modeling, such as image classification and object detection. However, the inherent computational complexity of ViTs―stemming from their quadratic self-attention mechanisms and massive parameter counts―poses significant challenges for deployment on resource-constrained edge devices. This paper proposes LightViT, a novel lightweight ViT architecture optimized for low-power edge scenarios. By integrating three key innovations―grouped self-attention, dynamic sparse connections, and hardware-aware mixed-precision quantization―Light ViT achieves a 75.2% Top-1 accuracy on ImageNet-1K with merely 4.1M parameters, while reducing FLOPs by 42% and power consumption by 55% compared to ViT-Tiny. Extensive evaluations on drone-captured datasets and real-world smart camera deployments demonstrate its robustness and efficiency. Furthermore, a systematic ablation study validates the contribution of each design component, and hardware-level optimizations on NVIDIA Jetson Nano showcase 23.4 FPS inference speeds under 3W power constraints. This work provides a comprehensive framework for bridging the gap between high-accuracy ViTs and practical edge applications, paving the way for scalable and sustainable edge AI solutions.