The Industrial Application and the Method of Improvement of Ganomaly
In recent years, deep learning technology, a well-developed image recognition method, has introduced to improve the quality of detection and adaptability to the diverse situation in the field of automatic optical inspection (AOI). However, the available data is highly imbalanced towards normality (i.e. a large amount of normal data and a small amount of abnormal data). It becomes a major challenge in the process of industrial application of deep learning; therefore, GANomaly, one of the best deep learning model of anomaly detection proposed in 2018, has drawn great attention from scientists. By learning the normal data, the generator of GANomaly can provide good quality fake images with a probability distribution which is similar to the input image. In addition, it has a discriminator to extract the image information. The residual score between the input image and the fake image created by the generator is a critical factor in anomaly detection. Nevertheless, the detection capabilities of GANomaly in industrial inspection have not been thoroughly discussed and verified. In this paper, the public industrial inspection dataset MVTec Anomaly Detection (MVTec AD) and wood samples from actual production lines were used to examine the performance of GANomaly. The situation of mixing abnormal data into the training data was fully investigated. Additionally, the area under the curves (AUCs) of sample inspection were improved by modifying different residual score to inspect industrial data.
Keywords - Deep Learning, Automatic Optical Inspection (AOI), GANomaly, Wood Inspection.