A Lightweight Track Feature Detection Algorithm Based on Element Multiplication and Extended Path Aggregation...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Existing track inspection methods face challenges with high computational load and insufficient real-time performance.
• Large model parameters hinder efficient deployment in resource-constrained environments.

Purpose of the Study:

• To develop a lightweight track feature detection module (YOLO-LWTD) to address computational and real-time performance issues.
• To enhance the efficiency and accuracy of track inspection for embedded systems.

Main Methods:

• Integration of the StarNet module into the backbone network to improve feature characterization.
• Utilizing a lightweight extended path aggregation network with C3K2-Light in the neck for efficient multi-scale feature fusion.
• Employing a lighter detection header (Detect-LADH) to reduce feature decoding complexity.

Main Results:

• The YOLO-LWTD model achieved a 0.5% increase in precision, 2.0% in recall, and 0.8% in mean average precision (MAP) compared to the benchmark.
• Inference speed improved by 38.1% to 163 FPS.
• Model size was compressed to 1.5 MB, a 71.1% lightweight rate.

Conclusions:

• The proposed YOLO-LWTD model offers an energy-efficient and lightweight solution for track detection.
• The model demonstrates superior performance in precision, recall, and speed, making it suitable for embedded deployment and real-time processing.