EXTRACTING TEXTUAL INFORMATION FROM BRIDGE INSPECTION RECORDS USING DEEP LEARNING

Abstract

To date, a large number of inspection records have been compiled following bridge inspections. However, most of these records have been saved either as hard copies or in PDF format, making it difficult to utilize the accumulated data. If data could be automatically extracted from these inspection records, and a database of information records created, it would be possible to analyze the data in a huge number of past inspection records. However, many structural element numbers which are important for obtaining information, such as the location of damaged members, overlap with the lines of structural members in bridge drawings, making it difficult to extract them through general optical character recognition (OCR) processing. This study extracted the element numbers in bridge drawings in the inspection records utilizing object-detection based on deep learning. The results confirm that it is possible to extract the element numbers that overlap with the straight lines with considerable accuracy. This paper is the English translation of the authors’ previous work [Yamane, T. et al.: , (2020). Extracting textual information from bridge inspection records using deep learning, Intelligence, Informatics and Infrastructure, Vol. 1, No. J1, 71-77 (in Japanese)].