Date of Award
Fall 11-25-2025
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Electrical and Computer Engineering
First Advisor
Richard Povinelli
Second Advisor
Cristinel Ababei
Third Advisor
Joseph Domblesky
Abstract
With the continuous growth in the need for artificial intelligence driven manufacturing, automatic detection of metal surface quality has become a key issue in forging production. Traditional manual detection methods are not only inefficient and subjective but also have low accuracy when analyzing workpieces with complex surfaces and diverse colors. To solve this problem, this dissertation proposes two methods for metal surface defect detection and classification. This dissertation begins with the background of forging and summarizes existing literature on industrial vision systems and surface defect detection methods. Basic image processing techniques, machine learning algorithms, and deep learning foundations—including artificial neural networks and convolutional architectures—are also reviewed to support the proposed methods. First, this dissertation proposes a metal surface defect detection method based on image reconstruction. By constructing an ideal defect-free reference image and comparing it with the actual image, accurate defect location is achieved in a complex background. Experimental results show that this method performs well in terms of Dice coefficient (0.92), Intersection over Union (0.86) and Hausdorff distance (3.40), which is significantly better than the traditional Sobel edge detection method (Dice is 0.45, IoU is 0.31, HD is 5.31), verifying the significant advantages of the proposed method in defect area identification and boundary accuracy. Second, this dissertation proposes a deep classification framework based on ensemble learning for classifying six different metal surface defects. This method integrates two convolutional neural networks as base learners and a support vector machine as the meta learner. The proposed method achieves accuracy, recall, precision, and F1 scores of 96.7%, 96.9%, 96.7%, and 96.6%, respectively. This is in comparison to a baseline method with accuracy, recall, precision, and F1 score of 97.0%, 91.6%, 91.1%, and 91.0%, respectively. In summary, the two methods proposed in this dissertation improve the automation and accuracy of metal surface defect recognition from the two levels of detection and classification and provide practical technical support for an intelligent quality inspection system in the forging industry.
Comments
Doctor of Philosophy (PhD)