.png){kind=icon}
.png){kind=icon}
Return
A Hardware Efficient Joint Compression Framework for Vision Transformers
.png){kind=icon}
OA
AI
DOI：10.1109/access.2026.3712528.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Vision Transformers achieve strong accuracy in image classification, yet their high computational cost and limited compatibility with integer-oriented execution graphs hinder edge deployment. Existing approaches typically focus on quantization or token sparsification in isolation, whereas the direct combination of these methods can produce unstable compression decisions because low-precision arithmetic distorts token saliency estimation. To address this issue, this paper presents HEViT, a deployment-oriented joint compression pipeline for DeiT-style Vision Transformers. The key design is a cross-precision decoupling mechanism that estimates token importance through an auxiliary full-precision control path during compression-aware training, while keeping the final inference graph aligned with INT8 execution. On top of this mechanism, HEViT integrates token merging and pruning, structured weight pruning, and integer-friendly nonlinear approximation to reduce sequence and parameter redundancy. Experiments on ImageNet-1K with DeiTs show that HEViT achieves up to 48% MAC reduction and a $1.56\times $ speedup over a matched INT8 baseline on Raspberry Pi 5, corresponding to a $9.6\times $ speedup over the FP32 baseline, while maintaining competitive accuracy.
Keywords:
Vision Transformer
edge deployment
INT8 quantization
joint compression
structured pruning
token sparsification