学会発表

基本情報

氏名 永田 寅臣
氏名(カナ) ナガタ フサオミ
氏名(英語) NAGATA Fusaomi

発表題目

Defect Detection Using Deep Convolutional Neural Networks, Support Vector Machines and Template Matching Techniques, (Keynote Speech),

代表発表者名

Fusaomi Nagata

共同発表者名

Kenta Tokuno, Keigo Watanabe, and Maki K. Habib

学会・会議名

2019 International Conference on Soft Computing & Machine Learning(SCML2019)

発表形態

口頭発表(招待・特別)

発表開始年月

2019/04/26

 

発表終了年月

2019/04/29

概要

Abstract of keynote speech : The authors already developed a user-friendly design and training application for deep convolutional neural networks (DCNNs) and support vector machines (SVMs) as shown in Fig. 1 [1, 2, 3]. The application allows students and novice engineers to easily design and train DCNNs and SVMs even if they are not familiar with the software development using C++ or Python. In this presentation, binary classification methods using DCNNs, SVMs and template matching techniques are introduced for visual inspection. DCNNs generally have several blocks consisting of convolutional, ReLU and pooling layers to accept image files (or feature maps) and produce more characterized ones to the following latter hidden layers, which lead to fully-connected layers and a softmax function layer for output. Two types of SVMs are firstly designed using the application, then they are trained using typical OK images without any defect in order to be able to distinguish images including defects from all images. It is assumed that the defects are crack, burr, protrusion, chipping, spot and fracture which appear in the manufacturing process of resin moulded articles. Figure 2 shows examples of images with the typical defects. Two types of pretrained DCNNs are respectively incorporated into the fore parts of two SVMs as feature extractors, in which compressed feature vectors extracted by the DCNNs are used as input vectors to the SVMs. The performance of the SVMs incorporated with the two types of DCNNs are compared and evaluated through training and classification experiments. In addition, a template matching technique is further applied to the SVM using AlexNet to narrow important target areas from original training and test images. This will be able to enhance the reliability and accuracy for binary classification using the SVM.