A Convolutional Neural Network for Arabic Document Analysis
Hassina Bouressace, János Csirik · 2019
In this study we apply a CNN (convolutional neural network) to segment a document image into its page components. Our approach consists of two main steps. Firstly, there is a new method of extracting layouts based on sharpness/smoothing filters, adaptive thresholding techniques, morphological operations, along with a CC (connected-component) technique and ARLSA (Adaptive run-length smoothing algorithm) for the patch extraction phase. Secondly, the extracted patches will be put into six classes (text, table, figure, title, legend, author) using a CNN, and four other classes (straight-line, text-line, block, article) using PP (projection profiles) analysis and geometric features. The method was tested on smartphone-captured newspaper images selected from the PATD [1] (Printed Arabic Text Database). The results indicate that the proposed method is suitable for detecting multiple labels in Arabic newspaper pages.