Optimization of deep learning-based faster R-CNN network for vehicle detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Hyperparameter optimization is vital for object detection model performance.
• Vehicle detection is a critical domain-specific application.

Purpose of the Study:

• To systematically optimize hyperparameters for the Faster R-CNN model in vehicle detection.
• To identify optimal configurations for enhanced detection efficiency and accuracy.

Main Methods:

• Evaluated base CNN architectures (VGG-16, ResNet-50, Inceptionv3).
• Assessed solvers (sgdm, rmsprop, adam), learning rates (10^-5 to 10^-3), and detection thresholds (0.1-0.3).
• Measured performance using average precision-recall (PR avg).

Main Results:

• Optimal performance (82% PR avg) achieved with ResNet-50, rmsprop solver, learning rate 10^-5, and threshold 0.1.
• Decreasing learning rate consistently improved network efficiency.
• Solver and detection threshold significantly impacted model performance.

Conclusions:

• Meticulous hyperparameter tuning is essential for improving object detection accuracy and reliability.
• The optimization methodology is applicable to diverse object detection tasks.
• Findings support enhanced surveillance, autonomous driving, and traffic management systems.