A Defect Detection Model for Industrial Products Based on Attention and Knowledge Distillation

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Industrial quality detection is crucial in manufacturing, leveraging technologies like big data, IoT, and edge computing.
• Existing deep learning object detection methods face challenges in specialized industrial scenarios.
• There is a need for robust industrial detection algorithms compatible with IoT and edge computing.

Purpose of the Study:

• To address the limitations of current deep learning methods in industrial quality detection.
• To propose novel isomorphic industrial detection models (T-model and S-model).
• To enhance feature fusion and detection accuracy for industrial defect identification.

Main Methods:

• Designed two isomorphic industrial detection models: T-model and S-model.
• Integrated swin-transformer with convolution in the backbone, incorporating a residual fusion path.
• Developed a dual attention module in the neck for improved feature fusion.
• Implemented a knowledge distiller for the lightweight S-model to boost detection accuracy.

Main Results:

• Experimental validation on four public industrial defect detection datasets.
• Demonstrated superior performance of the proposed models in industrial defect detection tasks.
• The T-model and S-model showed significant advantages over existing methods.

Conclusions:

• The developed T-model and S-model offer effective solutions for industrial quality detection.
• The integration of swin-transformer, attention mechanisms, and knowledge distillation enhances detection accuracy.
• The proposed models are well-suited for industrial applications, particularly in defect detection.