Machine learning based image analysis for surface defect inspection

• Main Author: Lee, Yong Xian
• Other Authors: Zheng Jianmin
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision; Engineering::Computer science and engineering::Computing methodologies::Pattern recognition
• Online Access: https://hdl.handle.net/10356/165900
• Institution: Nanyang Technological University

The progress in computer vision technology has significantly improved the reliability, effectiveness, and efficiency of defect detection. This is attributed to the availability of advanced optical illumination systems and appropriate image capturing devices that produce high-quality images. Additionally, deep learning, which is a crucial technology in visual inspection, has contributed to this progress. An approach called DifferNet [6] has been introduced, which estimates the density of feature descriptors extracted by a convolutional neural network (CNN) using normalizing flows. Although normalizing flow struggles with images that have high dimensionality, its key purpose is to handle data distributions with low dimensions. To overcome this challenge, a multiscale extractor can be employed, enabling normalizing flow to assign meaningful likelihoods that can identify defects through a scoring function. In this report, the objective is to explore how various pre-trained CNN models can help improve the features extracted for normalizing flows to be optimized further. To achieve that, different transfer learning models such as ResNet-50 [16], VGG16 [14], and EfficientNetV2 [28] trained on the ImageNet dataset [7] will replace AlexNet in the original paper. This report also explores the idea of utilizing the lower hierarchy of the pre-trained model to preserve localized nominal information loss mentioned in PatchCore [22].