Design And Implementation Of Image Classification Algorithm For Bank Bill

With the great development of information society,the degree of digitization of information processing in various fields is getting higher and higher,and traditional manual processing is gradually replaced by computer.In the field of bill processing,paper bill has been converted into image form for storage and processing.Bill image classification is an important link in bill processing and a necessary prerequisite for subsequent processing.However,in the actual business,there are a number of bill categories are very similar and only slight differences between categories.In this case,the bill image classification belongs to fine-grained image classification.Convolutional Neural Network model is the mainstream model to implement the fine-grained images classification,only category label used in model training is called weak supervision.Bilinear convolutional neural network performs well in fine-grained image classification based on weak supervision.This paper proposes a bill image classification algorithm based on the improved bilinear convolutional neural network model according to the characteristics of bill image.This paper first expounds the importance of bill image classification from many aspects such as the background and significance of bill image classification.This paper illustrates the relationship between bill image classification and fine-grained image classification by analyzing bill image data..Relevant basic theories are studied for bill image classification,including Convolutional Neural Network principle,VGGNet model principle,fine-grained image classification algorithm based on weak supervision,Global Average Pooling principle,PAC dimensionality reduction algorithm and image processing algorithm.This paper completes the design of bill image classification algorithm according to the characteristics of bill image,which includes bill image information extraction and the improved bilinear convolutional neural network model.The location set of ticket information is extracted,and the image block set corresponding to the information location in the image is captured as the input of the improved model.The improved bilinear convolutional neural network model is trained to complete classification of input bill image.This paper completes the design of bill image information extraction scheme.The specific steps include binarization,graying,denoising,tilt correction,frame detection and removal, and layout analysis.The reason of use and the effects of steps are described,the research and improvement of relevant algorithm is done,and the algorithm used in this paper to implement the processing steps is presented in each step.This paper completes the improvement of bilinear convolutional neural network.The input,structure and output of the model was improved respectively,including using the image block set of bill information as the input to improve the input,the fusion of two branch network to achieve computation partially share,and feature cross fusion between branche networks to improve the structure,at the end of each branch network adds Global Average Pooling layer and PCA dimensionality reduction to reduce the size dimension and the depth dimension of output feature,and at the last pooling layer adds the screening operation of output feature to improve the output.By using the bill image data for testing,the experimental results show that: the bill image information extraction scheme designed in this paper can accurately segment the bill layout and extract the bill information effectively.Compared the improved bilinear convolutional neural network model with the original bilinear convolutional neural network model,the classification accuracy and efficiency is improved slightly and the convergence rate of model is obviously accelerated.