Unsupervised Deep Learning for Handwritten Page Segmentation

Abstract

Segmenting handwritten document images into regions with homogeneous patterns is an important pre-processing step for many document images analysis tasks. Hand-labeling data to train a deep learning model for layout analysis requires significant human effort. In this paper, we present an unsupervised deep learning method for page segmentation, which revokes the need for annotated images. A siamese neural network is trained to differentiate between patches using their measurable properties such as number of foreground pixels, and average component height and width. The network is trained that spatially nearby patches are similar. The network's learned features are used for page segmentation, where patches are classified as main and side text based on the extracted features. We tested the method on a dataset of handwritten document images with quite complex layouts. Our experiments show that the proposed unsupervised method is as effective as typical supervised methods.

In-domain versus Out-of-domain transfer learning for document layout analysis

“…Finally, in [10] the authors tackle the challenge of limited ground truth availability by proposing an unsupervised deep learning approach for page segmentation. Their method involves the use of a Siamese neural network to differentiate between patches based on quantifiable properties, with a specific emphasis on the count of foreground pixels.…”

“…In the past few years, this problem has been tackled by various authors [9][10][11], who developed a set of few-shot-learning-oriented frameworks specifically aiming at leveraging the small amount of data available to generate more and more accurate predictions for the task at hand, producing results that are on par or even surpass previously available state-of-the-art approaches that relied on much more data. In the present paper, we tackle the problem from another point of view by exploring different transfer learning approaches as a way to make good use of alternative data sources to pre-train our models.…”

FAN-MCCD: Fast and Accurate Network for Multi-Scale Chinese Character Detection

“…Pertaining to historical documents, much research has been conducted recently for analyzing such documents. In this regard, the two kinds of approaches consist of conventional [22][23][24][25] and deep learning-like methods [26][27][28][29][30][31][32] to deal with detection of text in old documents. Phan et al [22] extracted characters depending on analyzing connected components.…”

“…Further, despite the use of sharing parameters to speed up the training, it is still not perfect enough for the character detection task because it still aches from the mislocalization problem. Ahmad et al [31] suggested a new page segmentation method that uses Siamese network to find the difference between patches; then, the extracted features were used to segment the page into main and side text regions, which means the authors handled the problem of pre-processing steps for document analysis without addressing the problem of word or character detection and recognition. In addition, expensive time was used for extracting the feature for every possible patch.…”

Handwritten Multi-Scale Chinese Character Detector with Blended Region Attention Features and Light-Weighted Learning

“…Recently, character-level detection, with its breakneck progress in the deep learning branch, has been handled as a feature extraction problem performed by a convolutional neural network CNN. In this regard, hierarchical, sequence-based, and segmentation-based models [ 1 , 2 , 3 ] have been presented to compensate for the lack of datasets, and pre- and post-processing approaches [ 4 ] provide pretty good solutions for precise detection tasks. However, it is worth mentioning that these methods may suffer from over-segmentation errors, and they might inaccurately position characters in a document.…”