Research On Chinese Bill Image Processing And Intelligent Recognition Technology

Because of its various font styles,huge character set and high character similarity,Chinese characters are more difficult to identify than traditional English letters,which brings great challenges to recognition research.Chinese-language bills are more challenging than ordinary standard Chinese documents because they are more complex and contain various seals,forms,and irregular patterns.Although there are various types of OCR equipment on the market,the recognition rate can reach 99.99%,but in reality,except for business cards,ID cards,etc.,the surface pattern has low complexity,small area,difficult to bend,and font structure.Compared with the target of a single category,it is easier to identify other thin paper objects similar to the bills.Currently,the OCR products on the market are not ideal for the recognition of bills with Chinese characters,and need to be further improved.This topic combines some traditional image processing algorithms and several deep neural network algorithms with better performance,and has made some improvements and optimizations to the traditional Chinese ticket identification system.This topic mainly does the following work:First of all,in the production part of the training sample,the method of augmenting the training data set not only adopts the traditional affine transformation and the method of adding random noise,but also introduces the conditional generation confrontation network,and makes a better expansion of the training samples.It is very beneficial to the training of the later recognition model.Secondly,the recognition model part adopts the deep convolutional neural network.This topic will make several neural network classification algorithm architectures with better performance at present.According to the actual needs of this topic,we have made structural adjustments.A better parametric model and use it in our final project.Second,to identify the deployment part of the model,we introduced a neural network acceleration bar.The scheme of deploying on the local PC first and then uploading the server is adopted,which replaces the scheme that the client is only responsible for uploading the bill image server for calculation.Finally,the error correction part of the identified result is inevitable in the actual situation due to uneven illumination or a small amount of damage to the picture of the bill.When the cutting is performed,the text will inevitably appear to be defective.Here we will have natural language.The processed algorithm is introduced into the error correction model,which is a combination of the LSTM network and the recognition model.