Ultra-Lightweight Fast Anomaly Detectors for Industrial Applications.

Area of Science:

• Industrial Engineering
• Computer Science
• Quality Control

Background:

• Quality inspection in pharmaceutical and food industries is vital for customer safety and brand perception.
• Visual inspection using machine vision and deep neural networks is common but struggles with diverse product formats.
• Existing AI solutions require extensive training and are less effective for rapidly changing production lines.

Purpose of the Study:

• To present a fast and adaptable image anomaly detection method for high-speed production lines.
• To address the limitations of current AI in quality inspection for variable product formats.
• To enable efficient defect detection with minimal computational resources.

Main Methods:

• Development of an ultra-lightweight algorithm for image anomaly detection.
• Focus on fast training capabilities suitable for diverse and frequently changing product formats.
• Implementation designed for real-time inference on high-speed production lines.

Main Results:

• The proposed method demonstrates ease and speed of training, even on low-power devices.
• Achieved inference times are suitable for real-time quality inspection scenarios.
• Successfully applied the algorithm to diverse real-world data from the food and pharmaceutical sectors.

Conclusions:

• The developed algorithm offers a flexible and efficient solution for visual quality inspection in dynamic production environments.
• It meets the requirements for fast training, real-time performance, and adaptability across different products and markets.
• This method enhances defect detection accuracy and efficiency in the pharmaceutical and food industries.