Deep learning-based digitalization of a part catalog book to generate part specification by a neutral reference data dictionary

Abstract

As the manufacturing sector enters the Industry 4.0 era, a higher level cooperative system must be established between manufacturers. Therefore, seamless sharing of information is required between companies, for instance, between an original equipment manufacturer and a parts manufacturer. However, books in PDF or image format that cannot be modified are still commonly used in the field to convey information. Moreover, locating the necessary information in documents, drafted based on unstructured data, is challenging. To overcome these drawbacks, this study proposes an end-to-end digitalization method to convert an image format catalog book into structured digital part specifications. The proposed method also defines a neutral reference data dictionary to ensure consistent digitalization to facilitate data interoperability, classifying catalog pages per part and identifying part numbers, detecting specification tables and recognizing texts in a table, and building part objects and their property objects from the texts extracted from the table. To validate our method, we conducted an experiment where catalog books for motor parts were digitalized. The experiment results exhibited excellent accuracy performance with 96.97% and 90.59%, in part object and property object conversion, respectively, when considering specifications. (c) 2022 Elsevier B.V. All rights reserved.