In recent years, with the continuous advancement of deep learning, Optical Character Recognition (OCR), including scene text detection and recognition technology, has developed rapidly. However, seals, an important element that can be seen everywhere in life and office, have not received attention, and the important topic of seal recognition is also faced with a variety of seal shapes, curved text content, complex background elements, and ubiquitous overlap. Text and many other problems. In the 17th ICDAR International Academic Competition held this year, a seal header recognition competition section was set up to seek solutions for targeted solutions to the pain points of seal header text detection and end-to-end seal header text content recognition.
ICDAR (International Conference on Document Analysis and Recognition) is a globally recognized authoritative academic conference in the field of document image analysis and recognition. It is also one of the earliest and most successful OCR events in the world. Important international competitions and standards related to the research progress of the latest technology of intelligent identification. The high technical difficulty and high practical applicability of the ICDAR series of events have attracted widespread attention from many scientific research institutions and technology companies. Since the event has been held, more than 7,821 teams from 119 countries have participated in it, and continue to inject impetus into the development of OCR technology. .
The ICDAR 2023 Reading the Seal Title of this competition was organized by a number of experts in the field. As a quasi-unicorn company in the AI field, Real Intelligence relies on its rich experience and leading position in the field of artificial intelligence for many years. Its technical solutions rank among the top ten together with Zhejiang University, Institute of Information Engineering, Chinese Academy of Sciences, China Mobile Cloud Center and other outstanding units!
The solution for this competition is the "two-stage" recognition method proposed by further innovation based on previous research results in the field: the first-stage detection scheme for curved text proposes to use a fifth-order Bezier curve to fit curved text, and the experimental It is proved that this improvement has achieved a better fitting effect; in the second stage, a more powerful backbone network was designed in the recognition of the header text, and an online enhancement strategy was designed for each seal type, and 100,000 pieces of synthetic data were synthesized for Training, and finally ranked in the top ten on Task 2. End-to-end Seal Title Recognition task. The excellent results in this competition once again demonstrate the outstanding full self-developed OCR capability of Real Intelligence.
At present, Real Intelligence has continued to iteratively optimize the OCR detection and recognition model based on product landing experience and collected corpus, and has realized its own functions such as general text recognition, form recognition, key value extraction, general certificate and seal recognition, and its AI The product IDP document review is an intelligent text processing intelligent product with cutting-edge technologies such as optical character recognition (OCR) and natural language processing (NLP) as the core, and is applied to business fields such as information digitization and risk warning to empower more Multi-industry digital upgrade and transformation brings scene solutions.