Neprižiūrimo dirbtinio intelekto autoenkoderinių modelių taikymo objektų vizualinių defektų aptikimui tyrimas / ; Investigation of the unsupervised artificial intelligence autoencoder models application for the object’s visual defects detection.

Currently, artificial neural network algorithms based on supervised learning are used in the field of engineering and manufacturing. They demand long and attentive labor for labeling data. During the last few years, in the experimental field of anomaly detection, more and more attention is given for unsupervised learning algorithms. Of which one the most popular – autoencoder (AE) was analyzed in this work. The aim of this study is to find the most effective AE for defect inspection and to determine how different modifications affect detections accuracy. In the first part of this work a short review is given about the most often employed deep learning autoencoders for the purpose of quality inspection. The second part describes selected MVTec AD database used in the training of unsupervised learning AE networks for object defect segmentation. Finally, after training various autoencoders with latent space, architectural and loss function differences on screw, hazelnut and bottle data, a study on accuracy was conducted with threshold independent characteristics (AUROC, AUPRC and AUIoU). The highest AUPRC score for detecting anomalies in screws was shown by AE with extended filter count – 29,9%, in hazelnut and in bottle was shown by context AE accordingly reaching - 51,9% and 41,4%. It was also observed that for different objects and different anomaly types, autoencoder performances were not persistent. AE models with additional regularization in loss function as well as context AE segmented high-contrast defects and reconstructed curved-edge objects more accurately while L2 loss function models performed better on asymmetrical image scenes segmenting screw defects more accurately.