Related Experiment Video
Updated: Aug 2, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
Laser-Visible Face Image Translation and Recognition Based on CycleGAN and Spectral Normalization
Mingyu Qin1, Youchen Fan2, Huichao Guo3
1Graduate School, Space Engineering University, Beijing 101416, China.
Sensors (Basel, Switzerland)
|April 13, 2023
Summary
This study introduces improved AI models for converting low-quality laser face images into clear visible images, enhancing nighttime face recognition accuracy significantly.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Biometrics
Background:
- Range-gated laser imaging captures faces in darkness but yields low-contrast, low-SNR images lacking color.
- These image quality issues hinder effective face observation and recognition, especially for long-distance applications at night.
Purpose of the Study:
- To develop advanced models for translating laser face images into high-quality visible face images.
- To improve the accuracy of face recognition using these enhanced images.
Main Methods:
- Proposed a Spectral Normalization-based Cycle Generative Adversarial Network (SN-CycleGAN) for laser-to-visible face image translation.
- Incorporated Y-channel content reconstruction loss to minimize mapping errors.
- Developed a feature retention-based face recognition model with triplet loss for domain constraint.
Main Results:
- The SN-CycleGAN model generated higher visual quality images with reduced errors and preserved structural features.
- Achieved a significantly lower Fréchet Inception Distance (FID) score (36.845) compared to existing models.
- The enhanced face recognition model achieved 76.9% Rank-1 accuracy, a 19.2% improvement over direct laser face recognition.
Conclusions:
- The proposed SN-CycleGAN effectively translates laser face images to visible ones, overcoming training difficulties and improving image quality.
- The feature retention face recognition model successfully utilizes translated images, demonstrating superior performance in nighttime biometric identification.
