Research On Block Diagram Recognition Technique Based On Digital Image Processing

With the development of computer technology and the popularity of smart phones and electronic reading devices,electronic documents have become one of the main ways for people to obtain information.Most printed documents are stored in the form of pictures.Thanks to the development of optical character recognition technology,people can directly extract the required data from the image for storage,processing and retrieval to reducing the burden of manual input.However,there are still a large number of block diagram images in the document.The existing OCR technology is difficult to extract their information directly.At present,the conventional block diagram recognition scheme mainly uses the traditional method to detect the contour and corner points of the block diagram,then relies on the manually defined features to extract the features to locate the key areas in the block diagram.But in practice there are many kinds of block diagrams,the situation may be very complex.There may be problems such as edge breaking,adhesion or other line element interference,which are difficult to deal with by traditional methods.The block diagram also has a complex two-dimensional structure,and the connection between elements must be considered in the identification.How to correctly analyze the structure of block diagram is also a difficult point in this field.This dissertation analyzes and studies several important problems in block diagram recognition: detection of key areas of block diagram,character segmentation and recognition in block diagram,structure recognition of block diagram by applying digital image processing and depth learning technology to block diagram recognition,and finally completes the block diagram recognition technology studied in this dissertation.How to label the collected data sets,how to enhance the data and how to reduce the noise,etc.The key regions of the block image are detected by using the YOLOv3 depth learning target detection algorithm,which effectively solves the problem that the traditional method is difficult to extract the features and the method is not universal.Based on the original YOLOv3 network,the multi-scale strategy is improved,and the recognition rate of multiple scales and arrow small targets is strengthened,which is better applied to the target detection of block image.In view of the fixed character font format in the printed block diagram image,this dissertation mainly uses the segmentation method to segment the characters,and puts forward a solution to the problem of adhesion characters.For character recognition,optimization for network structure based on Le Net-5 convolutional neural networks and obtains a character recognition model with high recognition accuracy.In terms of block diagram structure,this dissertation mainly analyzes the relationship between arrows,elements and connecting lines according to the results of block diagram region detection.The rules of block diagram structure identification are established and the logical relation of block diagram is determined under the constraints of this rule,which can effectively identify block diagram structure information.Finally,the characters of the block diagram are integrated and represented with the structure recognition results.