Study On Character Recognition Under Complex Background Based On Convolutional Neural Network And Multi-feature Fusion

With the rapid development of computer technology,more and more character information appears in the form of image.China has carried out extensive research on character recognition technology,and also promoted the development of database establishment,input automation,information printing and other technologies in the information age.Traditional character recognition technology is based on pattern recognition,and has certain defects in recognition efficiency,accuracy and intelligence.With the development of deep learning,character recognition technology has broken through the bottleneck of the traditional technical framework,as a new research hotspot,showing a broader application scenario.Although character recognition technology is advancing rapidly with the development of deep learning,there are still many problems for character recognition in complex background.In this thesis,the following researches have been done on the difficulty of acquiring character data with complex background and the difficulty of extracting character features under background interference:1)Constructed a complex background character data set and implemented the corresponding generation system.Neural network training requires sufficient data,but it is difficult to obtain sufficient character date with different complex backgrounds.In view of this problem,this thesis uses the background fusion method to generate complex background characters,and then uses data augmentation method to expand data and perform corresponding batch processing operations to generate data set.In order to facilitate the construction of data set,a data set generation system was designed and implemented,and the functions of each module were packaged and integrated,which improved the efficiency of constructing complex background character data sets.2)Designed a multi-feature fusion network model based on convolutional neural network.It is difficult to extract features from a single character image due to the interference of the background.To solve this problem,this thesis adds the features of the character image by fusing the feature information extracted by the character image using K-means and PCA in the deep network.Through this method,a richer feature map is obtained,thereby improving the recognition effect of characters in a complex background.3)Aiming at the recognition effect of multi-feature fusion network on characters with complex background,four optimization strategies of over-fitting,activation function,optimizer algorithm,and batch size are discussed in terms of the structure and parameters of the network model.The optimization methods are grouped to experiment,and the network model is adjusted through the analysis of the experimental results to obtain the optimal recognition effect.