Mold defects detection in painting using artificial intelligence / Mohamad Hilman Nordin

• Main Author: Mohamad Hilman , Nordin
• Format: Thesis
• Published: 2023
• Subjects: Q Science (General); TJ Mechanical engineering and machinery
• Online Access: http://studentsrepo.um.edu.my/15507/2/Mohamad_Hilman.pdf; http://studentsrepo.um.edu.my/15507/1/Mohamad_Hilman_Nordin.pdf; http://studentsrepo.um.edu.my/15507/

Defect detection is a crucial step in the restoration of damaged paintings. The process is usually lengthy and depends heavily on the qualitative visual judgement of an expert restorer. To enhance the efficiency and objectivity of this process, an artificial intelligence (AI)-based approach is proposed for mold defect detection in paintings. This study investigates the efficiency of four different techniques in identifying and classifying defects for mold defect detection that corresponds to rule-based system, machine learning and deep learning. The first technique examines four thresholding methods that convert grayscale images into binary images by statistically determining the threshold value. A new thresholding method called the Derivative Level Thresholding (DLT) is proposed and compared to existing binarization methods of Otsu’s Thresholding method, Minimum Error Thresholding (MET) and Contrast Adjusted Otsu’s Thresholding Method. The second technique utilizes Gray Level Co-Occurrence Matrix (GLCM) with Circular Proportionate Filter (CPF). The third technique is a machine learning classification that utilizes the feature extraction strength of the DLT. For this technique, eight machine learning models are trained and evaluated for mold classification: Logistic Regression (LR), Linear Discriminant Analysis (LDA), K-Nearest Neighbors (KNN), Classification and Regression Tree (CART), Gaussian Naïve Bayes (NB), Support Vector Machine (SVM), Random Forest (RF) and Extreme Gradient Boosting (XGB). Finally, the fourth method employed semantic segmentation using four different convolutional neural network (CNN) architectures to classify the sample images into mold regions and background. The CNN architectures evaluated are the Fully Convolutional Network (FCN), U-Net, SegNet, and a newly proposed Mold Segmentation Network (MSN). Experimental results demonstrate that the DLT method outperforms other thresholding methods in terms of precision, sensitivity, and accuracy. The pairwise GLCM with 40% CPF method achieves higher accuracy than the DLT with Morphological Filter for the classification task. Among the machine learning models, LDA exhibits the best performance with an accuracy of 93% accuracy and precision of 37%. For semantic segmentation, SegNet achieves the highest accuracy of 96% and precision of 66%. The proposed AI-based approach offers a promising solution for automated mold defect detection in paintings. It provides a more objective and efficient method compared to traditional techniques, which rely on the subjective assessment of conservators. The study contributes by demonstrating a novel application of AI by addressing the unique challenge of detecting mold defects, creating new datasets, and developing new AI techniques to complement art restoration process.