EnergyEfficient Transformer Compression for RealTime Semantic Segmentation on ResourceConstrained Edge Devices
Abstract
While Vision Transformers (ViTs) have redefined the state-of-the-art in semantic segmentation, their quadratic computational complexity O(N2) remains a barrier for edge deployment. We propose a multi-stage compression framework that integrates Dynamic Structural Pruning (DSP) and 8-bit Quantization-Aware Training (QAT). Our approach reduces the energy footprint by 65% while maintaining a Mean Intersection over Union (mIoU) within 1.2% of the baseline on the Cityscapes dataset. We demonstrate real-time inference (≥ 30FPS) on an NVIDIA Jetson Orin Nano and Raspberry Pi 5.
Fields of Study & Subjects
Advanced Neural Network Applications
IoT and Edge/Fog Computing
Domain Adaptation and Few-Shot Learning
Publication details
- Journal / source
- Journal of Emerging Technologies and Innovative Research
- Volume
- 13
- Issue
- 1
- Publication date
- 1 January 2026
- Language
- EN