Visual Anomaly Detection on Circular Plastic Parts Using Generative Adversarial Networks

In the recent years, automated quality control systems have been estabilished as the main method for anomaly detection, term which refers to the process of identifying and flagging any abnormality in the condition of the given components. Given their efficiency, many new methods were developed, mainly exploiting Computer Vision algorithms, but they have their limitations. In a similar way, many studies were applied on Neural Networks and Machine Learning algorithms, with the development of Convolutional Neural Networks, Transformers and Generative Adversarial Networks (GANs). The objective of this thesis is to develop an automated quality control system exploiting the generative and adversarial qualities of the current state-of-the-art methods based on Neural Networks. The main tool used for this task is the capability of the GANs to learn how a flawless input should look, so that the pipeline can identify inputs with anomalies. The developed solution was tested on a real world problem, aiming to indentify cracks and anomalies in plastic motor covers.