Document object classification method based on double-channel hybrid convolutional network

Abstract

The invention provides a document object classification method based on a double-channel hybrid convolutional network. The document object classification method is used for segmenting and classifyinglogic objects (texts, formulas, tables and images) in a document picture. According to the scheme, firstly, performing multi-mode matching recursion RLSA analysis on an input picture to determine segmentation coordinates; segmenting the input picture into different logic regions according to the segmentation coordinates; carrying out label marking, noise removal and category equalization processing on the region to obtain a classification data set; sending the two-dimensional image area piece to a two-dimensional CNN for training, extracting two-direction projections of the image, and sendingthe two-direction projections of the image to a one-dimensional CNN for training; and finally, using the first seven layers of the two convolutional networks as feature extractors, performing trainingof a final model through a two-channel hybrid classification network, and predicting the object category of the regional picture by using the model. According to the invention, the original two-dimensional picture and the projections in the two directions are respectively used as input, different characteristics are used, and the classification precision is improved.

Description

technical field [0001] The invention relates to the field of document object detection and recognition, in particular to a document object classification method based on a dual-path hybrid convolutional network. Background technique [0002] With the vigorous development of machine learning and deep learning in recent years, Document Image Understanding (DIU) technology has attracted more and more attention. Document picture understanding, as the name suggests, is to understand its content from document pictures. Document picture understanding can be specifically divided into steps such as page segmentation (also called region segmentation), region classification (also called block marking) and document object recognition, wherein the present invention corresponds to the first two steps, namely document object detection and recognition. [0003] The current page segmentation technology can be divided into two types in terms of steps. One is based on the pixel processing met...

AI Technical Summary

Problems solved by technology

In classification, two-dimensional AlexNet and one-dimensional AlexNet network with pixel projection as input are used as extractors, and input to a two-channel classification network for classification, which solves the shortcomings of two-dimensional networks that only focus on two-dimensional features and ignore one-dimensional features. , which improves the classification accuracy

Images