Surface defect classification in automotive coating process using Deep Learning

Machine and Deep Learning are two hot topics these days. Their performance levels have matched at some specific fields the human judgment or even surpassed it. However, what actually deep learning is mainly known for is for detecting completely or partially people, animals or everyday objects with high precision. However, when it comes to detect, identify or classify anomalies in the industry, the number of applications is not that big. The truth is that even when AI has shown an upgrade from classical Computer Vision approaches, the number of application that still use conventional cameras is very high. Which is why this project aims to reverse this situation, showing that Neural Networks have huge potential for quality inspection plant issues. From an inspection tunnel installed in the paint shop of an automotive plant, an update is proposed with this project. This tunnel uses classical Computer Vision techniques to identify all the possible defects in the body of each car. Lacking some information about the nature of every defect detected, a solution based on Neural Networks, specifically Convolutional, was proposed as the next step of the quality inspection process. Using the raw images obtained from the tunnel, the goal is to complete the actual information about the position and size of the defect with the corresponding label. However, the impossibility to access those images, made the project to take another direction by using another similar Image set, similar to the one studied from the original tunnel. From a public image set of defects on steels sheets, it has been build a solution that presumably could be exported to the original Dataset, due to the similarities between them. In this project it will be treated among other things the pre-processing and the augmentation techniques of the images, the most suitable net to be used for this task, the optimizer and the loss function and the best approach to the problem. Eventually, using the U-Net segmentation structure alongside a ResNet encoder, and a composition of 4 masks per images, it was able to classify defects with an accuracy of 93% using non defective images and 95% only using the defective images