Temporal-Quality Ensemble Technique for Handling Image Blur in Packaging Defect Inspection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning excels at surface defect inspection but struggles with packaging defects, particularly ingredient list verification.
• Existing methods primarily focus on high-quality images, neglecting defect detection in low-quality or blurred images.

Purpose of the Study:

• To develop an advanced inference technique for robust packaging defect inspection, specifically addressing challenges posed by low-quality images.
• To enhance the accuracy and reliability of defect detection in scenarios involving blurred or suboptimal image data.

Main Methods:

• Proposed a novel temporal-quality ensemble (TQE) inference technique combining temporal and quality weighting strategies.
• Temporal weighting considers image acquisition timing, while quality weighting prioritizes high-resolution, clear images.
• Evaluated TQE's general applicability using various convolutional neural networks (CNNs) like ResNet-34, EfficientNet, and ShuffleNetV2 as backbone networks.

Main Results:

• The TQE method significantly improved F1-scores in defect inspection tasks involving at least one low-quality image.
• Compared to single CNN models, TQE achieved 17.64% to 22.41% higher F1-scores.
• TQE outperformed average voting ensembles by 1.86% to 2.06% in F1-score.

Conclusions:

• The temporal-quality ensemble (TQE) offers a substantial improvement for packaging defect inspection, particularly when dealing with imperfect image data.
• TQE enhances both the accuracy and reliability of deep learning-based defect detection systems in real-world industrial settings.
• This technique provides a more robust solution for quality control, ensuring critical information like ingredient lists is accurately inspected even with image quality variations.