Research On Convolution Recursive Network Algorithm For Scene Text Recognition

Scene text recognition crosses two research fields,visual and language.It plays an important role in many tasks,such as scene understanding,auto-driving,certificate electronization,signature authentication and so on.However,the irregular distribution of text artwork itself and the degradation of image quality under complex acquisition conditions seriously affect the accuracy of text recognition.In view of this,this paper,based on convolution recursive network,studies the above two unresolved problems,which are as follows:1.Analyse the classical sequence text recognition algorithm and propose a three-stage architecture for text recognition tasks.Through a consistent process,the characteristics of each algorithm are studied,and the modular comparison and validity analysis are carried out to weaken the gap caused by different algorithm evaluation systems.Experiments on public datasets show that convolution recursive networks perform best in network recognition.2.To solve the problem that convolution recursive network has a low accuracy in recognizing deformed text images,this paper presents an irregular text recognition algorithm that fuses twin similarity measures.Unlike most irregular text correction algorithms which only change the network structure,this method analyzes the irregular samples that are misidentified and finds that the training sample data is not balanced.Therefore,the moving least squares distortion method is used to expand the training data and improve the data validity.On this basis,the training mode of the basic correction network is changed,and the twin structure is incorporated.By measuring the loss function of similarity,the distance between classes of similar samples is reduced,and the separation effect between classes of difficult samples is improved.Experiments show that this algorithm improves the performance of the basic correction network and can recognize heavily curved and distorted text.Compared with other irregular text recognition algorithms,it has the best average accuracy on multiple irregular text datasets and a more competitive text recognition performance on regular text datasets.3.To solve the low recognition rate of low-resolution text images by convolution recursive network,a low-quality text recognition algorithm combining gradient priori and attention perception is presented.Integrate the Super Resolution Quality Improvement module into the network,and in the image feature extraction phase,incorporate the sequence model to encode the sequence information of the text.By introducing a prior loss of the gradient contour,the sharpness of the image edge is enhanced,and the super-resolution image still conforms to the sequence and local edge consistency of the original image.In the text recognition module,a high resolution attention thermogram is used to monitor the recognition process of super-resolution images and to alleviate the attention drift during decoding.Experiments on an open super-resolution dataset show that the proposed method performs best on low-quality text images.On the visual effect level of image super-resolution reconstruction,PSNR and MSIM are used as evaluation systems,and good results are obtained.It is proved that improving the visual effect through super-resolution reconstruction can improve the accuracy of text recognition,and it provides a new way to solve the performance bottleneck of existing text recognition algorithms.