Deep Active Learning for Surface Defect Detection

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Current object detection models require large labeled datasets, limiting real-world deployment, especially in industrial defect detection.
• The high cost and effort of data labeling pose a significant challenge for complex industrial environments.
• Existing methods struggle with efficiency when dealing with limited labeled data for specialized tasks.

Purpose of the Study:

• To develop an active learning framework to reduce data labeling efforts in object detection for industrial defect detection.
• To enhance the efficiency and applicability of object detection systems in real-world industrial settings.
• To demonstrate that effective defect detection models can be trained with significantly fewer labeled samples.

Main Methods:

• Proposed an active learning framework incorporating Uncertainty Sampling to identify informative images for annotation.
• Developed an Average Margin method to dynamically determine the sampling scale for different defect categories.
• Implemented an iterative training and selection process to optimize the defect detection model.

Main Results:

• The active learning framework achieved the required performance levels with substantially reduced labeled data.
• Uncertainty Sampling effectively identified critical images, maximizing the learning gain from each annotation.
• The Average Margin method ensured balanced and efficient data selection across diverse defect types.

Conclusions:

• The proposed active learning framework significantly lowers the barrier to deploying object detection for industrial defect detection.
• This approach offers a practical solution for industries facing challenges with data acquisition and labeling costs.
• The method demonstrates the potential for training high-performance defect detection models with minimal labeled data.