Research On Lightweight Text Recognition Method Based On Binary Neural Network

With the rapid development of deep learning in recent years,many excellent deep neural network models have emerged.With the increasing data scale,the performance of these models has been continuously improved,marking that artificial intelligence has entered a new era.The deep neural network model is also widely used in OCR(Optical Character Recognition),which greatly improves the performance of text detection and recognition tasks,thus giving birth to more complex OCR application scenarios.At the same time,service carriers represented by mobile phones,electronic products and cloud services have accelerated the popularization of OCR and comprehensively promoted the accelerated landing and sustainable development of OCR technology industrialization.However,these complex deep neural network models need high storage capacity,and will cause a huge consumption of computing resources,which is difficult to meet the needs of fast and accurate practical applications.Therefore,how to design a smaller model that can run under the limited hardware constraints of mobile devices without affecting the accuracy is a key challenge.In order to solve this problem,this paper applies the binary neural network to character recognition,and limits the weight of the model to +1 and-1,then uses the full precision network model to distill the knowledge of the binary neural network to improve the recognition accuracy,and finally quantifies the model and deploys it to the Android system to realize the preliminary application of the binary neural network in character recognition.The main work of this paper is as follows:(1)A CRNN(Revolutionary Recurrent Neural Network)character recognition model based on binary neural network is proposed.Firstly,the binary neural network is used as the backbone network to extract image features,and then the features are input into the bidirectional LSTM(Long Short Term Memory)network to continue to extract the sequence features of characters.Finally,the input and output lengths are aligned through the CTC(Connectionist Temporary Classification)algorithm to calculate the loss function to obtain the final character recognition result.The model greatly reduces the demand for computing resources and improves the recognition speed.(2)A binary neural network knowledge distillation method based on FSP matrix(Flow of Solution Procedure)is proposed.The FSP matrix is obtained by calculating the inner product of the characteristic graphs in different networks,which is used to define the change process between different layers.Then the FSP matrix corresponding to the teacher network and the binary neural network is used L2-loss to calculate the loss function,so that the knowledge can be distilled from the teacher network to the binary neural network.This method can alleviate the information loss of binary neural network,help the binary neural network to extract more effective features,and improve the accuracy of character recognition.(3)The model deployment based on QAT(Quantification Aware Training)quantification method is designed and implemented.MNN(Mobile Neural Network),a deep learning reasoning framework,is used to get the optimal quantization model through the pseudo quantization process by using the quantization perception training method to reduce the model size and improve the reasoning performance.A simple and easy-to-use lightweight character recognition app is designed,and the character recognition function is realized on the mobile phone.