Extraction Technology of Legal Document Elements of Intellectual Property Cases Based on Convolutional Deep Belief Network
Year: 2024
Type: conference-paper
Abstract: Legal documents in intellectual property cases often contain multi-modal information, which is indispensable in retrieving and managing relevant materials, including texts, images, videos and other multi-modal resources. However, text recognition from multi-modal materials in natural scenes remains challenging due to complex backgrounds, image blurring, and irregular text distribution. Those problems are particularly common for images initially generated by humans compared with those synthesized by machines. To address those challenges, this paper focuses on developing a text recognition method for human-generated images. It proposes a novel method for text detection in natural scenes that integrates the Maximally Stable Extremal Regions (MSER) algorithm with a Convolutional Deep Belief Network (CDBN) to enhance text recognition accuracy. The proposed method is evaluated utilizing the ICDAR2011 and SVT datasets, with experiments conducted to inspect the impact of key factors, including hidden layer depth, the number of iterations, choice of pooling methods, and introduction of random noise on the performance of the CDBN architecture. The results show that increasing the number of hidden layers enables the network to extract more specific and detailed feature information, empowering deeper analysis of scene text images and more accurate characterization of textual features.
Author: Yineng Xiao
Institution: Nanjing University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Funder: Government of Jiangsu Province
OpenAlex ID: W4412524709