Research on Target Image Classification in Low-Light Night Vision.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Low-light imaging is crucial for military and civilian applications but suffers from low brightness, high noise, and loss of detail.
• Existing challenges hinder effective object detection and classification in ultra-low light environments.
• Night vision images present significant dimness and feature information loss.

Purpose of the Study:

• To enhance low-light night vision images for improved target detection and classification.
• To evaluate and compare image enhancement algorithms for low-light conditions.
• To develop and assess deep learning models for classifying targets in enhanced low-light images.

Main Methods:

• Image enhancement using Histogram Equalization (HE), Adaptive HE (AHE), and Contrast Limited AHE (CLAHE).
• Texture feature extraction using Gray-Level Co-occurrence Matrix (GLCM).
• Development of classification models using Backpropagation (BP) neural networks, VGG16, and ResNet50 convolutional neural networks.

Main Results:

• CLAHE was selected as the optimal enhancement algorithm based on peak signal-to-noise ratio and mean square error.
• A VGG16-based model achieved 92.1% classification accuracy for targets including vehicles, people, and license plates.
• The VGG16 model demonstrated superior performance, improving accuracy by 4.5% over BP and 2.3% over ResNet50.

Conclusions:

• CLAHE effectively enhances low-light images for subsequent analysis.
• Convolutional neural networks, particularly VGG16, combined with image enhancement, significantly improve low-light target classification.
• The developed approach offers a robust solution for object detection and classification in challenging night vision scenarios.