Analysis And Research On Binarization Problem Of Degraded Document Image

For the convenience of reading,the foreground text and background areas of most documents are purposely presented in high contrast.The binarization of images is a very important step-in most document analysis and recognition system.When it comes to degraded document images,the binarization processes become a challenging task.For general text images,binarization is very simple,but for degraded document images,they usually suffer from degradations such as uneven illumination,image contrast changes,character blur,background color leakage,infiltration,ink,paper aging,smudges,paper creases,etc.,and these degradations make the binarization processes more difficult.Furthermore,it is difficult to segment the foreground text and background because of the complexity of the image content and the different scales of the characters.The goal of document image binarization is to convert a given grayscale image or color document image into a black and white image.Due to the need for rapid retrieval and circulation of a large number of degraded documents,the binarization of degraded document images is a job that cannot be ignored.This paper mainly focus on the binarization of degraded document images.Firstly,the research background and current situation of the field are introduced.Secondly,the existing classic image binarization algorithms are described,including global threshold method,local threshold method,binarization algorithm based on statistical learning and deep learning-based binarization algorithm.And then,for the problem of binarization of degraded document images,we propose two binarization algorithms to obtain the binary form which not only,discard the noise and background information but also retain meaningful foreground text information.The first one is the local binarization algorithm with multi-threshold fusion,which is based on the traditional image processing method.We fuse the edge-based local binarization algorithm with the improved Sauvola algorithm to ensure the separation of foreground and background in the document image and improve the quality of image binarization.the other is the binarization algorithm based on cascading depth neural network.We first use a shallower network to extract feature maps at different scales using convolution operations.Then,the feature map obtained by deconvolution is combined with the feature map obtained by convolution at a unified scale to reconstruct the foreground image.After the shallow network structure,a deeper network is cascaded,and the final feature map of the shallow network is combined with the original image as the input of the deep network,and the obtained model can more effectively distinguish the background noise and the foreground,in order to optimize the final binarization result.Both algorithms have certain advantages in terms of background noise suppression and foreground text retention.At present,we have applied the degraded document image binarization algorithm based on cascaded deep neural network to the practical project.