Asymmetrical Contrastive Learning Network via Knowledge Distillation for No-Service Rail Surface Defect Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Materials Science

Background:

• Surface defect detection (SDD) is crucial in industrial applications.
• Existing deep learning models for SDD primarily use RGB data, neglecting valuable depth information.
• Current dual-stream models increase computational cost and parameter count.

Purpose of the Study:

• To propose a novel deep learning framework for trackless surface defect detection.
• To address the limitations of existing RGB-only models by incorporating depth features.
• To develop a parameter-efficient model that achieves high performance comparable to larger dual-stream networks.

Main Methods:

• A dual-stream teacher model (ACLNet-T) was developed to extract both RGB and depth features.
• A single-stream student model (ACLNet-S) was designed for parameter efficiency.
• Knowledge distillation techniques, including contrastive distillation, multiscale graph mapping, and attentional distillation, were employed to transfer knowledge from ACLNet-T to ACLNet-S.

Main Results:

• The proposed student model (ACLNet-S*) achieved performance comparable to the teacher model (ACLNet-T) with an eightfold reduction in parameter count.
• ACLNet-S* outperformed 16 state-of-the-art methods on the NEU RSDDS-AUG dataset.
• The model demonstrated strong generalization capabilities across three additional public datasets.

Conclusions:

• The proposed knowledge distillation approach effectively transfers multimodal features, enabling a compact yet powerful surface defect detection model.
• ACLNet-S* offers a promising solution for efficient and accurate industrial surface defect detection.
• The method highlights the potential of combining RGB and depth data with advanced distillation techniques for improved performance and efficiency.