Research On Deformed Chinese Document Correction Based On Deep Neural Network

Document image processing has a wide range of applications in the fields of office automation,digital libraries,and industrial automation.Compared with scanners,digital cameras are low in cost,easy to operate,and highly popular.When the surface of the document is bent,it is easy to cause image distortion to a certain extent,resulting in a decrease in the character recognition rate.Therefore,it is necessary to correct the document image to improve the recognition rate of characters in the document image.In this paper,a correction algorithm for estimating deformation parameters using deep neural network is proposed for Chinese printed document images with natural curved surface captured by handheld cameras.It is assumed that the book is in a flat state,and the polynomial deformation parameters of the book and camera pose parameters are simultaneously estimated by the deep neural network,thereby correcting the distortion image.The main work of this paper is as follows:(1)For the most common application scenarios,this paper mainly studied the curvature of the surface of the document caused by naturally open and horizontally placed books,and fit it with a cubic polynomial.Firstly,successive characters of Chinese characters were selected in the document,and the normalized coordinates of these characters were used as input of the deep neural network,and the corresponding deformation polynomial coefficients and camera pose parameters were output as the corresponding.Then,based on the perspective projection principle,the additional perturbed training samples were automatically generated.After training,the deep neural network can estimate the curved document deformation coefficient and the camera pose parameters with the normalized character coordinates.In this paper,the non-deformed document image and the deformed document image were studied separately,and the ideal results were obtained.(2)In order to determine the positions of the characters,the Most Stable Extreme Value Region algorithm was used to obtain the preliminary positioning result of the character,and then the non-maximum suppression algorithm was used to remove the overlapping detection result.Finally,the repetition of the single character is removed according to the character merging algorithm.Since the result of the detection does not include the row index information,the row index information in the preprocessing determines the operation,and the character coordinate values required for the input end of the deep neural network are automatically extracted.(3)Considering that the half-width and full-width characters in the printed body have different sizes,in order to correctly normalize the character coordinates,the support vector machine was used to classify the characters for characters with abnormally changed spacing,and the character spacing information was adjusted in combination with the character spacing information.For the blanks at the beginning and end of one paragraph,based on the consistency of neighborhood changes,a character completion algorithm was designed to obtain a sequence of characters that are completely full of consecutive rows.In this paper,different document image under different bending situations in the practical scene were collected with different camera pose,then character positioning and character preprocessing were performed,and the normalized coordinates are input into the deep neural network to obtain the deformation and camera pose parameters.The forward correction and reverse correction algorithms were used to correct them and compare with the classic "four-point method" correction results.The recognition rate of the OCR recognition software was used to evaluate the effect of the correction algorithm.The experimental results show that the proposed correction algorithm had a better correction effect on the plane document deformation image and the curved document deformation image.