| Abstract [eng] |
The methods analyzed in the work are designed to identify the surface defect. As human resources are currently insufficient, unreliable and do not pay off in the long run, automated systems are currently in place that detect a surface defect in a very short time, sometimes less than a second. Automated systems are usually implemented on the basis of neural network. These are artificial intelligence systems, neural and deep neural network systems. This paper reviews a GAN network based on a deep convolutional neural network, also called a DCGAN, and its ability to detect a surface defect. This task uses the DAGM2007 database, which has 10 different data sets that are divided into training and testing data. This database also has ground true images that are used to evaluate the results. The paper uses a discriminator model, which classifies the segmented surface area, and label it as good and as defective otherwise. The discriminator model was selected for each of the DAGM2007 datasets individually. The chosen discriminator model is modified from its initial structure and adapted to the particular data set to which it is assigned. The discriminator model is evaluated by accuracy, which describes how well the discriminator model is able to distinguish a surface defect from a surface which have no defect. After modifications of the discriminator model and application to the individual data set, 98.96% accuracy was obtained in identifying defects in the second data set. The lowest accuracy, 90.31%, was obtained by identifying defects in the fourth data set. |
