Research On Natural Scene Text Recognition Based On CRNN Algorithm

Computer vision has always been a research hotspot in the field of images.The objects in the images that contain textual information,such as road signs and billboards that may appear in natural scene images,can improve the image's contextual information and semantic information.Through the efforts of scientific researchers,many excellent natural scene text recognition methods have been proposed.These methods have gradually developed from the recognition of simple regular text to the recognition of irregular text,such as curved and perspective.However,in the face of complex natural scene conditions,how to improve the accuracy of text recognition algorithms and obtain better recognition results is still a research hotspot in the field of computer vision.This paper proposes a natural scene text recognition algorithm based on Convolutional Recurrent Neural Network(CRNN).The algorithm consists of a multidirectional irregular text recognition module and a semantic segmentation-based attention module.Combining the advantages of CRNN and semantic segmentation algorithm,the feature of multiple directions in the image is used to identify irregular text.The CRNN algorithm is a regular text recognition algorithm with high accuracy.The multidirectional irregular text recognition module obtained by improving the CRNN algorithm can recognize irregular text.This module first extracts the feature sequence of the image and the corresponding character position possibility sequence from four directions;then fuses the feature sequence of the four directions through the encoder to obtain the text feature sequence of the image;finally,the decoder predicts the final character sequence results.In order to further improve the accuracy of text recognition,this paper proposes an attention module based on semantic segmentation,which processes the image before the multidirectional irregular text recognition module.This module uses the semantic segmentation algorithm to perform semantic segmentation processing on the input image,and suppresses the background and noise areas in the image to improve the accuracy of subsequent text recognition modules.In order to verify the performance of the proposed CRNN-based natural scene text recognition algorithm,this paper tests on regular text datasets and irregular text datasets,respectively.The experimental results show that the proposed multidirectional irregular text recognition module achieves 97.3% accuracy on the regular text dataset SVT and 65.1% on the irregular text dataset CUTE80.Meanwhile,compared with most existing algorithm,the proposed multidirectional irregular text recognition module has less parameters.Combined with the semantic segmentation-based attention module,the multidirectional text recognition module achieves an accuracy rate of 66.2% on the irregular text dataset CUTE80.The experimental results prove that the CRNN-based natural scene text recognition algorithm can obtain better natural scene text recognition results.