Research Of Pattern Recognition Based On The Local Structure Topological Relationship Modeling

The image automatic classification and recognition are the hot research spot in the fields of pattern recognition and image processing.The icons and text in the image are the important research object to realize content-based image retrieval, image content analysis,image understanding and other application.This thesis studies a research of pattern recognition based on the local structure topological relationship modeling.From the perspective of human cognition,based on the biomimetic pattern recognition theory, we analysis that this new method is better closer to the function of human being based on the local structure and the global topological relationship.So we proposed a general framework of pattern recognition.In this framework,the methods of local structure feature extraction,topological relationship construction and modeling can be selected according to characteristics of the object.Based on the framework proposed above,we take the two-dimensional pattern image as object.The sliding window is used to convert the static image to feature sequence.The local structure feature is extracted in each sliding window.The topological relationship is constructed by the order of window in the sliding process.Based on the Artificial Neural Network(ANN) and Hidden Markov Model (HMM) the local and global topological relationship can be modeling.ANN has the features of strong learning capability,auto-adaptation,anti-noise and high speed.So it is used to model the local structure.HMM has the capability of dealing with time series.So it is used to model the global topological relationship.Finally,we implement a general method for pattern image recognition.In order to verify the system's effectiveness and versatility,we take a small scale icons images and a large scale chinese character image as experiment object.The result show that this method can be successfully applied to different types of object recognition.Meanwhile,it reflects a characteristics of human matter cognitive .This method has the capability of anti-jamming and anti-deformation.It shows a good recogniton result in the situation of incomplete image.In order to turn the research result into use,we developed a simple video caption recognition system.The caption are multiple font and multiple size chinese character.The system use the method above as recognizer,and add the function of caption detection,caption segmentation and character segmentation.The performance of the system meet the expected requirements.