Classifying and Characterizing Fast Nodular Iron Casting Metallographies by Applying a Similarity Search Method

Metallographic analyses of nodular iron casting methods are based on visual comparisons according to measuring standards. In fact, iron foundry workers have a poster that describes several characterizations of the metallographies and, showing the real metal in a microscope, they try to subjectively check the similarity between those examples and the real one. Currently, there are new approaches related to the application of machine vision and deep learning classifications. Although these aforementioned methods are more precise and accurate, they are more resource consuming, difficult to manage, and less scalable than other simpler methods that do not use the classical way of working with images. Moreover, for day-by-day work, this kind of precision is not needed, and this task must be carried out as fast as possible. 

Hence, this research work presents a novel approach to apply the same kind of comparisons carried out by human beings, but with the precision of a computer. Specifically, we construct a well-characterized vector database, populated with several metallographies analysed using accurate methods. Then, all images are represented by an embedding that tries to transform them into a vector representation to, finally, create the final classification and characterization of a specific metallography when applied a similarity search method in our learnt knowledge database.

Acknowledgement: This research has been performed within the “AZTERLAN-DEUSTO Technology Room” collaboration scheme.