Related Experiment Video
Updated: Jan 12, 2026
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
729
PUNet: a lightweight parallel U-Net architecture integrating Mamba-CNN for high-precision image segmentation
Zhaoyan Xie1, Xiaowei Li2, Hongyao Ma3
1Shandong Jiaotong University, Haitang Road 5001, Jinan, 250357, China. 215036@sdjtu.edu.cn.
Scientific Reports
|November 6, 2025
Summary
PUNet, a novel lightweight neural network, achieves real-time image segmentation on edge devices. It efficiently captures local and global features using depthwise separable convolutions and Mamba modules, enabling high-precision results with minimal parameters.
Area of Science:
- Computer Vision
- Deep Learning
- Edge Computing
Background:
- Real-time, high-precision image segmentation on resource-constrained devices is difficult.
- Traditional convolutional networks struggle with long-range dependencies and high computational costs.
Purpose of the Study:
- To introduce PUNet, a lightweight parallel U-Net variant for efficient image segmentation.
- To enable real-time inference on mobile and edge devices.
Main Methods:
- PUNet integrates depthwise separable convolutions (DSConv) and a structured state-space Mamba module in a dual-path encoder.
- A squeeze-and-excitation skip connection adaptively fuses multi-scale features.
- The model has only 0.26 million parameters and linear computational complexity.
Main Results:
- PUNet achieved a validation Dice score of 0.9208 on CamVid and 0.7902 on CRACK500.
- Mean Intersection-over-Union scores were 0.8643 (CamVid) and 0.6612 (CRACK500).
- Outperformed other lightweight models significantly.
Conclusions:
- PUNet offers a computationally efficient solution for real-time, high-precision image segmentation.
- Its architecture is well-suited for resource-constrained mobile and edge platforms.