| At present,document image is becoming more and more important part of electronic information transmission,but in the process of collection,processing and transmission of document image may cause various degrees of quality loss,reduce the efficiency of information transmission,so the quality assessment of document image can provide an important reference for the efficiency of information transmission.In practical applications,there is often no reference image,so this thesis mainly adopts the assessment method without reference to evaluate the quality of document image,and objectively evaluates the document image from three aspects that have a high impact on image quality,including clarity,distortion degree and writing quality.The main contents are as follows:1.When evaluating the distortion degree of document image,aiming at the reasons of the distortion,a method is proposed to evaluate the distortion by using the tilt and deformation degree of text line: Firstly,the point set at the bottom of text line is obtained by image morphological processing method; Then,using the idea of least squares to fit the line of the point set,the skew degree of the current text line is obtained,and the average skew degree of the whole text is obtained.At the same time,judging whether each text line is deformed according to the mean value,and then getting the overall deformation degree.Finally,a comprehensive assessment of the distortion degree is obtained according to the influence of the two factors on the distortion.2.In order to evaluate the definition of document image,aiming at the interference of light and shadow on the selection of threshold value in the process of binarization,and the influence of invalid information points(mostly seen in background patterns in checks)on the statistics of edge information before and after second order blurring,combined with the idea of second order blurring and the morphological processing flow of image,a new assessment method of second order blurring is proposed: First,the original image and the image processed by Sobel operator are binarized,and then the two are combined to get the final binarized result.The binarization process after fuzzy operation of the original image is the same as above.Finally,the changes of the image contents before and after the blur were counted,and the final objective assessment results were obtained.Compared with the traditional feature extraction method,this method can lock the change of information before and after the fuzzy in the content that people are interested in,so the assessment result is closer to the subjective assessment.3.In the assessment of handwriting quality,aiming at the feature that the legibility of document image content is the key to information transmission,an assessment method is proposed to reflect the overall handwriting quality from the mean value of single character legibility.Firstly,the adhesion of the refined skeleton was divided into single point adhesion,single line adhesion and complex adhesion.Single point adhesion was treated by projection method and single line adhesion was treated by drip algorithm.For complex adhesions,density clustering was used to segment loose adhesions,and K-Means clustering was used to treat compact adhesions.Then,according to the position relation between connected domains and the morphological characteristics of the side radicals,the characters are combined into multiple parts.For the segmented characters,the existing relatively mature CNN network is used for training and recognition; Finally,the accuracy of segmentation algorithm,network recognition ability and character recognition accuracy are combined to get the assessment of handwriting quality. |
