GDB: Gated convolutions-based Document Binarization

Abstract

Document binarization is a key pre-processing step for many document analysis tasks. However, existing methods can not extract stroke edges finely, mainly due to the fairtreatment nature of vanilla convolutions and the extraction of stroke edges without adequate supervision by boundaryrelated information. In this paper, we formulate text extraction as the learning of gating values and propose an endto-end gated convolutions-based network (GDB) to solve the problem of imprecise stroke edge extraction. The gated convolutions are applied to selectively extract the features of strokes with different attention. Our proposed framework consists of two stages. Firstly, a coarse sub-network with an extra edge branch is trained to get more precise feature maps by feeding a priori mask and edge. Secondly, a refinement sub-network is cascaded to refine the output of the first stage by gated convolutions based on the sharp edge. For global information, GDB also contains a multiscale operation to combine local and global features. We conduct comprehensive experiments on ten Document Image Binarization Contest (DIBCO) datasets from 2009 to 2019. Experimental results show that our proposed methods outperform the state-of-the-art methods in terms of all metrics on average and achieve top ranking on six benchmark datasets.

An Efficient Transformer-CNN Network for Document Image Binarization

“…But it can also be clearly seen that the results of document image binarization based on deep learning are far better than those of traditional methods, so the algorithms of document image binarization in the past three years are basically studied on the technology of deep learning. Such as literature [57][58][59][60][61][62][63][66][67][68][69]72,73].…”

“…The models of document image binarization based on deep learning technology are mainly divided into two research directions. One is the model obtained based on Convolutional Neural Network (CNN) [19,[47][48][49][50][51][52][53][54][55][56][57][58][59][60][61][62][63], and the other is based on Generative Adversarial Networks (GANs) [66][67][68][69].…”

“…In summary, the above CNN-based models [19,47,50,[57][58][59][60][61] generally have the advantages of CNN, simple structure design, fewer parameters, and high computational efficiency, especially friendliness to limited computing resources. While the models [64][65][66][67][68][69][70][71] operating within the GAN framework have the characteristics of GANs.…”

“…In recent years, the achievements in document image binarization research have primarily been realized through deep learning [19,[47][48][49][50][51][52][53][54][55][56][57][58][59][60][61][62][63]. Neural networks have demonstrated remarkable ability to segment foreground text and background of document images, as demonstrated by a comparison of the deep learning algorithm [63] with the traditional method [46] (winner of DIBCO2018) for binarization of handwritten document images, as shown in Figure 3.…”

An Efficient Transformer–CNN Network for Document Image Binarization

“…But it can also be clearly seen that the results of document image binarization based on deep learning are far better than those of traditional methods, so the algorithms for document image binarization that have been developed over the past three years have generally been based on deep learning technology. Such as literature [57][58][59][60][61][62][63][66][67][68][69][72][73][74], especially the model of Biswas et al [72] has almost obtained the best results on most of the DIBCO datasets. Biswas et al's model is a document binarization encoder-decoder architecture based on a Tokens-to-token vision transformer, employing a progressive tokenization technique to capture the local information from an image to achieve more effective binarization results.…”