Related Experiment Video
Updated: Jul 21, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
EfficientSegNet: Lightweight Semantic Segmentation with Multi-Scale Feature Fusion and Boundary Enhancement
Le Zhang1, Mengwei Li1, Peng Zhang1
1School of Instrument and Electronics, North University of China, Taiyuan 030051, China.
Sensors (Basel, Switzerland)
|October 16, 2025
Summary
EfficientSegNet offers a lightweight solution for semantic segmentation, improving accuracy and efficiency on resource-constrained devices. This new model addresses challenges in multi-scale targets and blurred boundaries for computer vision applications.
Area of Science:
- Computer Vision
- Deep Learning
- Image Segmentation
Background:
- High-precision semantic segmentation models are computationally intensive and memory-demanding, limiting their use in embedded systems.
- Traditional methods struggle with multi-scale targets and object boundaries due to loss of spatial information during deep feature extraction.
Purpose of the Study:
- To develop a lightweight and efficient semantic segmentation network.
- To address limitations in handling multi-scale targets and object boundaries.
- To enable efficient deployment on resource-constrained devices.
Main Methods:
- Proposed EfficientSegNet architecture.
- Integration of Cascade-Attention Dense Field (CADF) module.
- Integration of Dynamic Weighting Feature Fusion (DWF) module.
Main Results:
- EfficientSegNet achieves a balance between segmentation accuracy and computational efficiency.
- Demonstrated robust performance on multiple public datasets.
- Reduced computational resource requirements while preserving global semantic information and local detail.
Conclusions:
- EfficientSegNet provides a viable solution for real-time semantic segmentation on resource-constrained devices.
- The proposed architecture effectively handles multi-scale targets and object boundaries.
- The network offers improved segmentation accuracy with reduced computational overhead.
