Quality Control of PET Bottles Caps with Dedicated Image Calibration and Deep Neural Networks.

Area of Science:

• Industrial automation
• Machine vision systems
• Deep learning applications

Background:

• Product quality control is crucial in manufacturing, especially for food and pharmaceuticals, demanding precise analysis.
• Machine vision and image processing offer contactless solutions for production line monitoring.
• Deep neural networks excel in image analysis but face challenges with inference and training times in high-throughput industrial settings.

Purpose of the Study:

• To introduce a novel data preprocessing method for industrial quality control.
• To enable the use of lightweight Convolutional Neural Networks (CNNs) for efficient product analysis.
• To reduce prediction and training times without compromising accuracy.

Main Methods:

• Development of a new data preprocessing technique leveraging prior knowledge of the optical system.
• Implementation of a lightweight CNN model for quality control of polyethylene terephthalate (PET) bottle caps.
• Comparative analysis of the proposed method against standard models on ImageNet.

Main Results:

• Achieved at least a five-fold reduction in prediction and training time.
• Maintained accuracy comparable to standard, heavier models.
• Demonstrated the effectiveness of the preprocessing method in conjunction with lightweight CNNs.

Conclusions:

• The novel preprocessing method significantly accelerates CNN-based quality control.
• Lightweight CNNs, when combined with this preprocessing, are viable for high-performance production lines.
• This approach offers an efficient solution for contactless product quality inspection.