Nghiên cứu ứng dụng thuật toán nhận dạng cấu trúc bảng dựa trên phát hiện đối tượng

0 Người đánh giá. Xếp hạng trung bình 0

Tác giả: Hồng Phan Hải, Đại Dương Ngô, Lê Phương Phạm

Ngôn ngữ: Vie

Ký hiệu phân loại:

Thông tin xuất bản: Khí tượng thủy văn 2021

Mô tả vật lý: 45303

Bộ sưu tập: Metadata

ID: 409068

Thêm vào giỏ Liên kết toàn văn

Tóm tắt
Chủ đề

Nhận dạng cấu trúc bảng là vấn đề quan trọng trong số hóa tài liệu. Với sự phát triển của các kỹ thuật học sâu hiện nay việc phát hiện bảng đã có nhiều bước tiến lớn, trong khi đó nhận dạng cấu trúc bảng vẫn gặp rất nhiều khó khăn do cấu trúc bảng phức tạp, đặc biệt với dữ liệu thực tế. Bài báo này đề xuất một phương pháp ứng dụng mô hình Cascade mask R-CNN x101FPN deconv để nhận dạng hàng và cột. Bài báo cũng đề xuất sử dụng mô hình Faster R-CNN để nhận dạng các ô trong bảng, từ đó đưa ra cấu trúc bảng. Thuật toán đề xuất được đánh giá trên tập dữ liệu phổ biến như TabStructDB và các tài liệu thu thập được từ các trạm Khí tượng thủy văn. Kết quả thực nghiệm đạt 90% độ chính xác trên các tập dữ liệu này. Thuật toán có khả năng áp dụng hiệu quả vào việc nhận dạng cấu trúc bảng của các tài liệu thông thường
đặc biệt thuật toán có khả năng xử lý với các tài liệu lịch sử và các chữ viết tay, phù hợp với đặc điểm tài liệu của ngành Khí tượng thủy văn. Từ đó góp phần vào việc số hóa tài liệu, lưu trữ và truy xuất thông tin dữ liệu của ngành Khí tượng thủy văn., Tóm tắt tiếng anh, Table structure identification is an important issue in document digitization. With the development of current deep learning techniques, the detection of tables has made great strides, while table structure identification still faces many difficulties due to complex table structure, especially with real data practice. This article proposes a method to apply the Cascade mask model R-CNN x101FPN deconv to identify rows and columns. The paper also proposes to use the Faster R-CNN model to identify the cells in the table, thereby giving out the table structure. The proposed algorithm is evaluated on popular datasets such as TabStructDB and documents collected from Hydrometeorology stations. The experimental results reached 90% accuracy on these datasets. The algorithm is capable of effectively applying to the identification of the table structure of common documents
especially, the algorithm is capable of dealing with historical documents and handwriting, in accordance with the document characteristics of the hydrometeorology industry. Since then, it contributes to the digitization of documents, storage and data retrieval of hydrometeorology industry.

Tạo bộ sưu tập với mã QR