Improving Visual Defect Detection and Localization in Industrial Thermal Images Using Autoencoders.

Area of Science:

• Industrial Engineering
• Computer Vision
• Machine Learning

Background:

• Reliable anomaly detection in thermal images is vital for industrial defect detection.
• Challenges arise from smooth transitions in image sequences, leading to contaminated training data.
• Autoencoder-based methods struggle with slightly contaminated datasets and threshold determination.

Purpose of the Study:

• To improve anomaly detection reliability in thermal image datasets.
• To address challenges posed by contaminated training data in industrial defect detection.
• To enhance threshold determination for classifying healthy and defective industrial products.

Main Methods:

• Utilized autoencoder models trained on healthy thermal images with Mean Squared Error (MSE) and Structural Similarity Index Measure (SSIM) loss functions.
• Applied three anomaly scores for classification: MSE, SSIM, and Kernel Density Estimation (KDE).
• Introduced MSE+ and SSIM+ methods, including SSIM-based quantitative anomaly localization parameters.

Main Results:

• Achieved average accuracies: MSE (95.33%), SSIM (88.37%), and KDE (92.81%) on thermal image datasets.
• Demonstrated that combining anomaly scores improves classification accuracy.
• Showcased improved performance with KDE, particularly for contaminated healthy training data.

Conclusions:

• Combining anomaly scores effectively separates healthy and defective data, overcoming challenges in threshold determination.
• The proposed method enhances anomaly detection reliability in industrial thermal imaging.
• Kernel Density Estimation (KDE) proves beneficial when training data is contaminated.