Deep Learning-Based Detection of One and Two-Column Textual Blocks in Camera-Captured Pashto Documents Images

The paper explores the layout analysis and classification task of Pashto document images, a field with limited research due to the language’s low-resource status. It uses Document Image Analysis (DIA) to detect one-column and two-column text blocks from Pashto documents captured using handheld cameras. A novel dataset containing real-world documents is annotated using bounding boxes to distinguish between one-column and two-column layouts. Layout analysis plays a vital role in Document Image Analysis (DIA) and answers questions about the nature of Document Images integral parts (components). These components are mainly text, graphics, and tabular data. The detection/identification of such elements is initially dependent on the general layout of a document. Textual columns are one of the basic layout structures present in document images. Such textual columns need identification as one-column or two-column before feeding the textual blocks into an Optical Character Recognition (OCR) system. Also, it becomes more crucial if the document images belong to a low-resource language like Pashto. Thus, this work contributes mainly in two aspects. First, it contributes to creating a dataset containing camera-captured document images with one-column and two-column patterns. Second, it proposes a deep learning model to identify/detect one-column and two-column textual blocks in the Pashto text images. The proposed model utilizes a reputed variant of a Faster Region-based Convolutional Neural Network (R-CNN) called single shot detector (SSD). The evaluation was done by examining the test set and a mean average precision (mAP) of 89% is achieved as a baseline.