A text extraction framework of financial report in traditional format with OpenCV

Due to intensified off-balance sheet disclosure by regulatory authorities, financial reports now contain a substantial amount of information beyond the financial statements. Consequently, the length of footnotes in financial reports exceeds that of the financial statements. This poses a novel challenge for regulators and users of financial reports in efficiently managing this information. Financial reports, with their clear structure, encompass abundant structured information applicable to information extraction, automatic summarization, and information retrieval. Extracting headings and paragraph content from financial reports enables the acquisition of the annual report text’s framework. This paper focuses on extracting the structural framework of annual report texts and introduces an OpenCV-based method for text framework extraction using computer vision. The proposed method employs morphological image dilation to distinguish headings from the main body of the text. Moreover, this paper combines the proposed method with a traditional, rule-based extraction method that exploits the characteristic features of numbers and symbols at the beginning of headings. This combination results in an optimized framework extraction method, producing a more concise text framework.