Machine learning (ML) has significantly changed industrial image processing in recent years. Classical image recognition algorithms are often replaced by convolutional neural networks (CNNs). These are often more accurate and robust. However, the question remains as to which method of image recognition is most appropriate for each individual application. Mainly performance and computational efficiency are relevant questions. For the further development of CNNs, it is also interesting to understand more precisely how different CNNs learn and represent features. To answer these questions, a classical image recognition algorithm is compared with three CNNs of the VGG16 architecture in course of the elaboration. The networks are trained according to three different learning paradigms: as an autoencoder, as a classification model and as a deep embedded clustering (DEC) model. Images of an automatic inspection system for uranium pellets are used for the comparison. The system is located at Advanced Nuclear Fuels GmbH. CNNs achieve almost as good classification results as the classical algorithm. The highest accuracy is achieved by the classification model (0.996), closely followed by the autoencoder (0.994). However, none of the CNNs is able to achieve the required false positive rate of 0.000. In terms of robustness, a changed color scheme does not affect any of the methods. In contrast to the classical algorithm, the CNNs are also hardly affected by a distorted perspective. Within the CNNs, the DEC model and autoencoder are more robust against a distorted perspective than the classifier. Regarding training efficiency of CNNs, it turns out that the training process for the classification model and autoencoder takes nearly the same amount of time. For the DEC model, the training process takes ˜12.5 times as long. Feature representation in the CNNs are not very similar even in the first layer. For later layers of the networks, the similarity decreases. The representations of the DEC model and classification model tend to be the most similar and those of autoencoder and DEC model least similar. The results of the thesis highlight the potential and limitations of CNNs in industrial image recognition. Nevertheless, further research and improvement is necessary to meet accuracy requirements in critical applications and enhance explainability of the CNNs.
Date of Award | 2024 |
---|
Original language | English |
---|
Supervisor | Kurt Niel (Supervisor) |
---|
Comparision of Classical Fuel Pellet Inspection with Supervised and Unsupervised CNNs
Wiggerthale, J. (Author). 2024
Student thesis: Master's Thesis