STMS-YOLOv5: A Lightweight Algorithm for Gear Surface Defect Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Materials Science

Background:

• Deep learning object detection models face challenges in gear surface defect detection due to high computational demands and complex architectures, leading to suboptimal speed and accuracy.
• Existing methods struggle to balance efficiency and performance in real-time industrial inspection scenarios.

Purpose of the Study:

• To propose a novel lightweight deep learning model, STMS-YOLOv5, for efficient and accurate gear surface defect detection.
• To address the limitations of current algorithms in terms of speed, accuracy, and computational cost.

Main Methods:

• Implemented a lightweight backbone using the ShuffleNetv2 module to minimize GFLOPs and parameters.
• Integrated transposed convolution upsampling to enhance network learning capabilities.
• Embedded the max efficient channel attention mechanism to counteract accuracy loss from the lightweight backbone.
• Utilized SIOU_Loss for bounding box regression to accelerate model convergence.

Main Results:

• Achieved high inference speeds of 130.4 FPS (gear dataset) and 133.5 FPS (NEU-DET steel dataset).
• Reduced model parameters by 44.4% and GFLOPs by 50.31% compared to baseline models.
• Obtained high mean Average Precision (mAP@0.5) of 98.6% on the gear dataset and 73.5% on the NEU-DET dataset.

Conclusions:

• The proposed STMS-YOLOv5 model offers a significant advancement in lightweight deep learning for industrial surface defect detection.
• Demonstrated superior performance in terms of speed, parameter reduction, and accuracy, validating its effectiveness and generalization capabilities.