Research And Implementation Of Bill Recognition System Based On Deep Learning

With the continuous development of deep learning and social intelligence,automatic recognition of bill text has become the mainstream research direction.Through deep learning,the text recognition of bills is led to the intelligent track,which can get rid of the cumbersome manual review process and effectively save the cost of human and financial resources.This thesis takes VAT bills,train tickets,transportation tickets as research objects,and conducts research on text recognition algorithms based on computer vision and deep learning.The main work includes the following four parts:1.Due to the dense text of VAT bills,when locating the text of inclined VAT bills,the positioning accuracy is low,and the traditional tilt correction algorithm generally has the problems of low operation efficiency and poor anti-interference ability.Therefore,this thesis proposes a VAT bill correction algorithm of Unet++ that integrates the attention mechanism,and selects eight network models to train and test the network on the homemade datasets.The experimental results show that under the premise that the correction angle is floating within±0.2 degrees,the correction effect of the Unet++network model fused with the attention mechanism is the best.2.In order to adapt to the characteristics of the dense text of the bill and the small character target,a text localization algorithm based on improved YOLOV5 is proposed.Optimize and improve the data augmentation algorithm,feature fusion mode and prediction box screening strategy of the traditional YOLOV5 network,and enhance the sensitivity of the YOLOV5 network to small targets.The text detection model is obtained by training the homemade datasets of text positioning,and the test experiment is carried out on the practical bill image.The results show that the algorithm has high positioning accuracy.3.In order to identify text regions more accurately,a text recognition algorithm of CRNN with convolutional attention mechanism is proposed.By improving and optimizing the basic convolution unit,introducing a deep convolution module,an attention mechanism and a deep residual module,the depth of the network is expanded,the sensitivity of the CRNN network to text features is improved,and the network is trained on the homemade datasets of text recognition.The experimental results show that the improved CRNN network proposed in this thesis has higher text recognition accuracy and better recognition effect on VAT bills,train tickets and traffic tickets than the traditional CRNN network.4.In order to facilitate the promotion and application of the researched algorithm,this thesis develops a bill recognition system,the system includes a login module,a bill upload and recognition module,a history query module,and an exit module,and the clustered deployment architecture is used to complete the system online.Relying on superior computer hardware and clustered deployment architecture,the system has high concurrency and throughput,ensuring the robustness and high availability of the system.