Laser curve extraction of a train wheelset based on an encoder-decoder network.

Area of Science:

• Computer Vision
• Machine Learning
• Railway Engineering

Background:

• Accurate segmentation of train wheelset laser curves is crucial for detecting defects and ensuring railway safety.
• Existing methods struggle with the rich local and simple semantic features present in train wheelset laser curve images.

Purpose of the Study:

• To propose an efficient encoder-decoder neural network for precise laser curve segmentation of train wheelsets.
• To design a network that effectively utilizes dense connections and upsampling modules for enhanced feature propagation and reuse.

Main Methods:

• Development of a shallow, high-resolution encoder-decoder neural network architecture.
• Implementation of dense connection mechanisms and upsampling modules to improve feature extraction and multi-scale context understanding.
• Training and evaluation on a dedicated train wheelset laser curve dataset.

Main Results:

• The proposed encoder-decoder network demonstrated superior performance compared to other neural networks for laser curve extraction.
• Achieved high scores: mIOU of 86.5%, Recall of 89.2%, Accuracy of 99.9%, and F1_score of 85.0% on the dataset.
• The network effectively reduced the impact of noise on laser fringe extraction.

Conclusions:

• The encoder-decoder network provides accurate and robust laser curve segmentation for train wheelsets.
• This method has significant potential for application in enhancing railway safety inspection systems.
• The designed network offers efficient feature extraction with fewer parameters and improved detail preservation.