Research Of Layout Structure-based Document Image Retrieval

With the development of multimedia technology and information age, various kinds ofimage information have been grown quickly; electronic documents have been widely used in allworks of life instead of paper forms of text images. With the size of image database increasing,the demand of information inquires to people are diversified, so how to find images peopleinterested in from image data base quickly in order to reduce cost, save time and improve thedegree of automation has become interested subject in the field of retrieval. The traditionalimage retrieval technology has been no longer content with the demand of users, thus a fastimage retrieval technology has important research value and extensive application. The retrievaltechnology of the commonly used contains two methods: text-based retrieval technology andfeature-based technology which has been wide application technology so far. The text image isthe special image that mainly contains text information as well as diagrams, and it is difficult todescribe with color and texture, so the key technology of text image retrieval is how to extractfeature and calculate similarity of the feature.Layout structure analysis plays an important role in feature extraction and image retrieval,based on the structure of layout and combined global properties as well as local characteristicsthat are used as index entry, a new text image retrieval method was proposed in the paper basedon existing algorithms. After feature extraction which’s described as vector and similaritycomparison, an algorithm based on spectral clustering was used to retrieve document imageswhich users are interested in, so the size of the data comparison was greatly reduced to someextent, which was beneficial to retrieve document image in a short time and improve theefficiency of retrieval.Firstly, the preprocessing of document images which includes noise removing, binarizationand Hough transform was done. Median filter method was used to remove isolated noise points,and then skew detection and correction was done using gradient difference combined withHough transforms, the paper used Bernsen and Ostu algorithm to binary image before theprocess of skew detection. Secondly, global and local features were detected was done usingBottom-up and Up-bottom method of layout structure. Effective area was located before featuredetection, according to finding maximum blank area, column information was extracted. Inaddition, the segmentation process of text and non-text area was segmented by MaximumGradient Difference, and paragraph feature was extracted by means of connected domain merger;Key block feature was extracted in the non-texted. Thirdly, the paper used Gaussion distancefunction to compare the distance of feature vector in the vector space model.A fast retrieval algorithm of document image was proposed in the paper, the algorithm wasfirstly divided into several classes, and further divided was completed by the thought ofclustering, which could reduce comparison time when users queried images. After finding thebiggest similarity class, the query image would be compared with each image in the class above. As a result, image candidate set was got in the last step. The result shows that the time ofretrieval was greatly reduced and retrieval efficiency was greatly improved at the same time ofensuring the accuracy of retrieval.