| As an important pre-processing step in the document analysis and recognition system,binarization can not only improve the visual quality of document image,but also have an important effect on subsequent detection and recognition.Traditional binarization algorithms are usually based on the artificial features and have poor effect on documents containing various types of degradation.Binarization algorithms based on deep learning use Fully Convolutional Networks for End-to-End segmentation of document images and obtain good results.However,Fully Convolutional Networks essentially learn the color difference of pixels,which make it easy to misclassify the texts with light color or the background noises with dark color.Moreover,the limitation of input size of Convolutional Neural Networks cause it to deal with the problem of uneven text distribution poorly.To solve the above-mentioned problems,this paper firstly introduce the attention-gated module which can enhance the text pixels and suppress the background pixels so as to increase the difference between the texts and noises and make the network outputs more accurate binary results.Based on this,a two-stage network is proposed to solve the problem of uneven text distribution.The first-stage network is used to generate the global and local prediction of the image.The second-stage network fuses the information of the two predictions and outputs accurate binary results with more clean background and more complete stroke structures.In this paper,a series of experiments are conducted on DIBCO datasets which are the benchmark datasets of document image binarization.In the ablation experiment,the binary performance of the network is improved after introducing the attention-gated module,which proves the module is effective.Compared with the single network,the performance of the two-stage network is obviously improved on the DIBCO2013 datasets in which the problem of uneven text distribution is major,which proves that the two-stage network can effectively solve this problem.In the comparison experiments,the proposed algorithms are evaluated on three datasets including DIBCO2017,DIBCO2018 and DIBCO2019,and then compared with other binarization algorithms.The results show that the binarization performance of the proposed algorithm is better than most binarization algorithms,and the highest score is obtained in the evaluation of DIBCO2018 datasets.In a word,this paper proposes some beneficial improvements to solve the problems of Fully Convolutional Networks in document image binarization task,and the experiments show that these improvements are effective. |
