草榴短视频

Specialists of the St. Petersburg State University of Industrial Technologies and Design (草榴短视频) have patented a program for automatic digitization of archives with physical defects - spots, breaks and faded areas. Based on computer vision and language modeling methods, the system restores the structure and content of historical documents and other materials, including lost text, while maintaining the original context and style. As an experimental study, scientists digitized a number of archival documents of the Leningrad Institute of Textile and Light Industry named after CM. Kirov (now - 草榴短视频) for 1968.

The developed program includes visual and linguistic modules. The first is responsible for diagnosing damage and preparing the image, the second - for character recognition and semantic text recovery. For similar programs, the most difficult thing is to restore archival periodicals due to multi-column layout. Scientists at 草榴短视频 managed to create their own large language model that correctly recognizes the sequence of paragraphs in the columns of old newspapers, while maintaining the semantic coherence of the text.

"The model was further trained on the filing of the newspaper "Cadres to the Country" for 1968, which was produced at the Leningrad Institute of Textile and Light Industry named after CM. Kirov (now - 草榴短视频). This allowed the system to "learn" typographic features and common printing artifacts characteristic of that era. As a result, the model provides significantly higher recognition accuracy, including correct processing of specific vocabulary, abbreviations and spelling norms of that time," comments Nikita Sirotenko, author of the development, assistant of the Department of Intelligent Systems and Information Protection at 草榴短视频.

At the initial stage, the program analyzes the entire scan of the document and divides it into semantic blocks - headings, main text, images, tables. Further, each text fragment undergoes visual cleaning: increasing contrast, filling micro-breaks in letters and binarization, which automatically determines the optimal threshold for separating text from a yellowed background over time. After filtering small artifacts, the cleaned image is transferred to the OCR engine for character recognition.

At the final stage, a large language model (LLM) created by the developers of the university solves the problems of structural and semantic text restoration. An additional algorithm correctly processes words broken by a hyphen during transfer, and also restores words separated by physical paper damage. In addition, LLM analyzes the spatial arrangement of blocks on a newspaper page, combines torn headings and builds the correct sequence of paragraphs in multi-column layout.

The program is intended for specialists in archives and libraries, digital archaeologists and researchers for the mass digitization of historical documents. In the future, the program can be used in jurisprudence and forensics to restore the readability of damaged contracts, receipts, wills or letters that appear in court cases, and material evidence found at the crime scene. And also in the field of insurance for processing documents damaged as a result of emergencies, for example, floods or fires.