8 From August 22nd to 24th, the Vision and Learning Young Scholars Seminar (VALSE 2022) sponsored by China Artificial Intelligence Society was held at Tianjin National Convention and Exhibition Center . Kingsoft Office was invited to attend the conference and discussed the cutting-edge technological achievements of computer vision and machine learning with top domestic scholars.
How to transform non-editable tables into electronic documents that are easy to copy, paste and edit has become a hot research topic and technical problem in both academia and industry in recent years. At the "Interpretable Deep Learning Algorithms and Vision Applications" sub-forum, Kingsoft Office Technology Director and CV Team Leader Xiong Longfei delivered a speech on "Table Restoration in Complex Scenarios", demonstrating the company's breakthrough progress in this field in recent years. .
Xiong Longfei pointed out that tables in standard PDFs or scanned and clear tables are relatively easy to identify, and many conversion engines support the conversion and restoration of such tables. But for tables in more complex scenarios, such as table files that are severely deformed, have complex layout, have random orientations, blurred pixels, have interference from seals or other background patterns, have nested relationships, or even have no table lines, most of the table files currently on the market The recognition and conversion effects of products are unsatisfactory.
In recent years, Kingsoft Office has reached the industry-leading level in the fields of OCR recognition and table recognition supported by CV technology. To solve the technical problem of table restoration in complex scenarios, Kingsoft Office has developed a recognition system. Through a variety of self-developed algorithms, it has achieved better results in interfering document elements such as dense text, irregular tables or frameless tables, seals, signatures, etc. Good recognition and conversion effects. In addition to meeting basic document content recognition,
has added deep learning capabilities and made a series of algorithm optimizations for text attributes, which can more accurately restore the document style, including color, font, bold, italics and underline formats. It can carry out fine recognition, significantly improve the degree of document restoration, and save users a lot of time and energy in manual adjustments in the later stage.
According to reports, the system consists of nearly 20 deep learning models, more than 100 algorithm modules, and more than 500,000 lines of code. It has complex functional modules and scheduling processes, and supports all types of image input, whether scanned or online. Downloaded images, cell phone screenshots, random photos, etc. can all be converted and recognized. It can be deployed on both the cloud and local ends, and supports the output of files in multiple formats such as xlsx tables, word documents, and html. At the
lecture, Xiong Longfei demonstrated the recognition capabilities of the system through a series of samples. For blurred, stamped form images, and even physical forms that were crumpled into a ball and expanded, they were all recognized by the system to a very high degree of restoration. Spend.
“This system can save a lot of time costs caused by manual editing in daily office, financial form conversion, government work and other fields, and greatly improve office efficiency.” Xiong Longfei said.
Kingsoft Office has persisted in deeply cultivating the office field for the past 34 years, adhering to technology to establish its business, and constantly polishing technology and product services. The company undertakes the research and development tasks of a number of major national scientific research projects, including the "Nuclear High-tech" major science and technology project, the Electronic Development Fund, the "863" high-tech research and development plan, etc., and has twice won the second prize of the National Science and Technology Progress Award.
company has continued to increase investment in research and development in the field of AI in recent years, and has listed OCR and layout recognition and understanding as key tasks. As early as 2018, the company established a CV team to focus on the R&D and application of image and visual projects on WPS and other products. It empowers various office scenarios through CV capabilities. Currently, it has been engaged in OCR text recognition, document correction, intelligent cutout, More than 20 functions such as font recognition, table restoration, layout restoration, and PDF editing have been launched with related products and services.
Currently, Kingsoft Office provides office services to users from more than 220 countries and regions around the world. As of June 2022, the company's main products have 570 million monthly active devices, including 232 million monthly active devices for the WPS office PC version and 334 million monthly active devices for the mobile version.
VALSE 2022 is hosted by China Image and Graphics Society and China Artificial Intelligence Society, Tianjin University is hosted by Tianjin University of Technology , Tianjin University of Science and Technology and Tianjin Artificial Intelligence Society and other units, aiming to provide computer vision , image processing, pattern recognition, and machine learning research fields to provide a stage for in-depth academic exchanges for Chinese young scholars. Up to now, VALSE has been successfully held for 11 times and is the largest machine vision academic event in China.
