Deep Learningを応用した不良品検出の基礎研究

抄録

Neural networks (NNs) with a hierarchical structure of four or more layers are called deep NNs (DNNs) and are recognized as a promising machine learning technique. In this paper, deep convolutional NNs (DCNNs) were designed to inspect defects such as crack, burr, protrusion and chipping which occurred in the manufacturing process of resin molded articles. An image generator was first presented to efficiently produce many similar images. Then, the DCNNs were trained with a large amount of training images of each category that the image generator produced. It has been confirmed that the trained DCNNs have the ability to classify sample images of the training test set into five categories of “OK”, “Crack”, “Burr”, “Protrusion”, “Chipping” with high recognition rates.