Related Experiment Video
Updated: Jun 21, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
515
An Improved U-Net Infrared Small Target Detection Algorithm Based on Multi-Scale Feature Decomposition and Fusion and
Xiangsuo Fan1,2, Wentao Ding1, Xuyang Li1
1School of Automation, Guangxi University of Science and Technology, Liuzhou 545006, China.
Sensors (Basel, Switzerland)
|July 13, 2024
Summary
This study introduces MST-UNet, an enhanced U-Net model for infrared small target detection. It improves segmentation accuracy by minimizing feature loss and enhancing feature extraction using multi-scale fusion and attention mechanisms.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Signal Processing
Background:
- Infrared small target detection is vital for military, security, and medical applications.
- Deep learning, particularly convolutional neural networks, shows promise but struggles with small targets due to scale, weak signals, and background noise.
- Existing methods often suffer from feature leakage and misdetection in infrared small target segmentation.
Purpose of the Study:
- To propose an enhanced U-Net model, MST-UNet, for improved infrared small target segmentation.
- To address the limitations of traditional convolutional neural networks in detecting small targets with weak signals and complex backgrounds.
- To enhance feature utilization and reduce information loss during the segmentation process.
Main Methods:
- Implemented MST-UNet, combining multi-scale feature decomposition, fusion, and attention mechanisms.
- Replaced maximum pooling with Haar wavelet transform for downsampling to preserve feature information.
- Introduced multi-scale residual units for enhanced contextual information extraction and feature expression.
- Integrated a triple attention mechanism to improve feature recovery and multidimensional information utilization.
Main Results:
- The MST-UNet model demonstrated significant improvements in target contour accuracy and segmentation precision.
- Achieved Intersection over Union (IoU) of 80.09% and normalized IoU (nIoU) of 80.19% on the NUDT-SIRST dataset.
- Effectively mitigated issues of feature leakage and misdetection common in infrared small target segmentation.
Conclusions:
- MST-UNet offers a robust solution for infrared small target detection and segmentation.
- The proposed method effectively enhances feature extraction and utilization, leading to superior segmentation performance.
- The integration of wavelet transform and attention mechanisms provides a promising direction for future research in this domain.
