Automatic Detection of Defects in Pickled Sheets

Manual identification of defects in products in a company can lead to loss of time, resources and money. This paper presents the development of an algorithm that, from a photograph taken of the defect in a product (pickled sheet) of a metallurgical company, the defect type is identified. This algorithm is based on the concatenation of two Machine Learning models: the first uses a feed-forward neural network, trained with tags obtained from a digital image
processing algorithm; the second is pre-trained using a Densenet121 convolutional neural network which is fed by the raw photographs of the defects. The results obtained are compared using separately the image classification model and the tag classification model, and finally the concatenation of both models. According to the results obtained in this work, it is confirmed that the use of digital image processing and Machine Learning techniques is a promising solution for the automation of this type of problems.