LWViT-Lightweight Vision Transformer for Edge Deployment via Adaptive Multi-Dimensional Pruning

The complex environments of construction sites present dual challenges for visual algorithms: the need for real-time performance and high accuracy. However, deploying traditional Vision Transformers (ViTs) on edge devices is hindered by substantial computational demands. To address this, we propose LWViT (Lightweight ViT), an adaptive model compression framework based on multi-dimensional task correlation. LWViT analyzes the interplay between the structural dimensions of ViTs and task-specific features to develop tailored pruning strategies. Experiments conducted on three benchmark datasets show that, at a pruning ratio of 0.80, LWViT achieves a 1.2× increase in inference speed while improving accuracy by 2.05% compared to the original ViT, outperforming state-of-the-art pruning approaches. This work offers an effective solution for the lightweight deployment of visual algorithms in edge computing scenarios.