Improvements of Classification Accuracy of Film Defects by Using GPU-accelerated Image Processing and Machine Learning Frameworks

Research on image classification for natural images are quite actively worked on and recent achievements using deep learning techniques are tremendous. On the other hand, image classification techniques of defects in industrial products are mostly kept secret, partly because defective images contain very sensitive information about the products and the confidential manufacturing technologies. With the help of a leading company in a visual inspection of film defects, we investigated the effectivity of using machine learning techniques for classification of defect images. We also made use of GPU to accelerate both image processing to assist detection of defects and machine learning. We propose the combination of deep neural networks with random forest classifier for image classification of film defects, which performed better than using either of the two techniques alone.