Related Experiment Video
Updated: May 11, 2026
12:08
From Voxels to Knowledge: A Practical Guide to the Segmentation of Complex Electron Microscopy 3D-Data
Published on: August 13, 2014
24.5K
Design and Implementation of a Lightweight and Energy-Efficient Semantic Segmentation Accelerator for Embedded
1School of Integrated Circuits, Southeast University, Nanjing 211189, China.
Micromachines
|March 27, 2025
Summary
Optimized ENet semantic segmentation achieves real-time performance on embedded systems. Structural simplification and 8-bit quantization reduce complexity while maintaining accuracy, enabling efficient deployment for autonomous driving and medical imaging.
Area of Science:
- Computer Vision
- Embedded Systems Engineering
- Deep Learning
Background:
- Real-time semantic segmentation is crucial for autonomous driving and medical imaging.
- Resource-constrained embedded platforms present challenges for efficient deep learning deployment.
- Efficient Network (ENet) is a lightweight model with low computational complexity.
Purpose of the Study:
- To optimize the ENet semantic segmentation network for enhanced efficiency and hardware compatibility.
- To reduce computational complexity through structural simplification and 8-bit quantization.
- To improve on-chip data storage and data transfer for embedded deployment.
Main Methods:
- Structural simplification of the ENet network.
- 8-bit quantization for reduced precision and computational load.
- Optimization of transposed convolution and dilated convolution computational units.
- Enhanced on-chip data storage and data transfer mechanisms.
- Deployment on Xilinx ZYNQ ZCU104 FPGA.
Main Results:
- Maintained 51.18% mean Intersection over Union (mIoU) accuracy.
- Achieved a frame rate of 130.75 FPS, meeting real-time requirements.
- Demonstrated low power consumption (3.479 W) and high throughput (460.8 GOPS).
- Reached an energy efficiency of 132.2 GOPS/W.
Conclusions:
- The optimization and deployment strategies effectively balance computational efficiency and accuracy.
- The optimized ENet is well-suited for real-time semantic segmentation on resource-constrained embedded platforms.
- The system meets the demanding requirements of applications like autonomous driving and medical imaging.
