E-Commerce Picture Text Detection And Recognition Based On Deep Learning

E-commerce has developed rapidly since its appearance in China in 1999.By the end of 2017,China's e-commerce market transaction volume reached 29.16 trillion yuan,including which the online retail market transaction volume reached 7 trillion,accounting for 50% of the global online sales share.China has become one of the largest and fastest growing countries in e-commerce.People can use the mobile terminal device to perform a few simple operations,and they can purchase the goods they need on the e-commerce platform.The traditional e-commerce management method is limited to text,mainly for text processing methods such as blocking and keyword filtering.Today,e-commerce platforms are increasingly using pictures to display information,which brings technological challenges to e-commerce management.Therefore,automatically reading the text information in the product information picture has become a hot issue.For text detection,this thesis implements a FCN-based text detection model.The method first extracts the features of the input image through ResNet50,and then fuses each layer of different sized features map which obtained by ResNet50,then achieve regression and classification.The removal for the steps of extraction,filtering and fusion of candidate regions improve the efficiency of the model.The predicted text boxes are combined by improved NMS to obtain the final test results.At the same time,in order to solve the problem of unbalanced data and accelerate the convergence of the model,the model introduces a loss function combined with the cross entropy loss of DiceLoss and the balance between instances.For text recognition,this thesis implements a model of sequence-based text recognition.The method includes four models: DenseNet feature extraction layer,spatial transformation layer,bidirectional LSTM sequence learning layer and transcription layer based on connected time series classification model.In the text recognition task,the training data has a significant effect on the model.This thesis first enhances the data to increase the robustness of the model.In order to extract more feature information,the model uses a dense network DenseNet as the feature extraction layer of the convolutional neural network.Spatial transformation and alignment of data through spatial transformation networks solve the problem of distorted and skewed text,then cope with extracted features by BLSTM.Finally,for the problem that the length of text in the image is different and the text is not easy to segment,the joint time classification model commonly used in speech recognition is used to transcribe the output prediction into the final result.