Research On Text Detection And Recognition In Natural Scene Images Based On EAST And ASTER

With the advent of the Internet era,the acquisition and storage of natural scene pictures is becoming more and more convenient.Natural scene pictures record people's lives and important information.The text in the picture contains rich semantic information.How to accurately extract the text from the natural scene image has important application value in the fields of video retrieval,license plate recognition,navigation system,and industrial production automation.Different from the document image,the background of the natural scene image is complicated,the sharpness of the picture is greatly different,and the rules of the text area are diversified,making the traditional OCR technology unable to apply to text recognition in the natural scene.In recent years,the technology based on deep learning has penetrated into various application scenarios.In this paper,the existing popular EAST text detection algorithm and ASTER text recognition technology are used to study the extraction of text in pictures of natural scenes.First,to solve the problem of insufficient receptive field capability,this paper implements an improved EAST text detection model.This model uses deep residual networks to capture deeper image features.Features are output with steps 8 and 16 before feature fusion.After the mapping layer,the ASPP module is added to expand the receptive field.Then,the feature fusion is performed by bilinear upsampling to the appropriate size and the shallow feature mapping in series,and finally output to the output layer through convolution.In network training,this paper modified the loss function to train the network by combining the two major loss functions,Focal Loss and dice loss.The test on experimental data proves that the model can well perform the text detection function in natural scenes.Then,in view of the difficulty of training the classic ASTER model,this paper implements an improved text recognition model.The model includes two modules: a rectifier network and a recognition network.The rectifier network retains the STN structure based on the STN structure for skewed and curved text."Correction" function;the recognition network uses a shared CNN and LSTM coding layer,combined with the CTC and attention mechanism decoding layer.Experimental results show that the network can reduce the difficulty of training the model without affecting the accuracy of text recognition.