Research On Key Technologies Of Text Line Analysis Based On New CNN Instance Segmentation Algorithm

The automatic processing of document image can greatly reduce human labor.As a crucial part of document analysis and recognition system such as optical character recognition and key words retrieval system,text line segmentation of document image is the key to improve the accuracy of character recognition,and it is of great significance for the digital storage of document image.It becomes non-trivial for unconstrained handwritten documents due to the existence of touching lines,multi-skewed lines and different writing styles.Therefore,text line segmentation of unconstrained document image is still a valuable research issue.In recent years,deep learning method has achieved outstanding achievements in research directions such as image segmentation.Image segmentation methods based on convolutional neural network have sprung up.Two methods based on DCNN are proposed in this thesis to solve the problem of text line segmentation.Combined with post process,the proposed methods can effectively deal with the problems of text line segmentation,such as adhesion,skew and different character sizes.The main works of this thesis are as follow:Firstly,a method based on semantic segmentation is proposed for text line segmentation.The document image is divided into two types:the main body area and the background area.The VGG16 net is selected as the backbone,and in order to build a fully convolutional network for text line segmentation,it is improved by adding batch normalization and dilated convolution.The output probability map of the network is binarized to get the main body area of different text lines,and then the CCS are grouped according to the nearest center line to get the text line segmentation result.A run length coding based method is employed to deal with the adhesion characters.Secondly,a method based on instance segmentation is proposed for text line segmentation.The DeepLab v3+network based on RseNet50 is built for instance segmentation.Each pixel of the input image is mapped to a point in n-dimensional feature space through the convolutional neural network.Then make the embedding vector close to the expected distribution by minimizing the discriminative loss function which is used for instance segmentation.The mean-shift algorithm is used to cluster the high-dimensional vectors to get a coarse segmentation result,and the post-processing based on thinning algorithm is used to fine-tune the segmentation resultsThe two proposed methods are used to experiment on the public handwriting segmentation dataset,and good segmentation results are obtained,which demonstrate the effectiveness of the proposed method.