| As an important preprocessing module of many practical application systems,scene text detection and recognition has received extensive attention in the field of computer vision.Many excellent text detection and recognition algorithms have been proposed and achieved remarkable results.However,the detection and recognition of irregular text in natural scenes still face great challenges,and the existing methods cannot achieve good results.Is still an open problem.Therefore,aiming at irregular text in natural scenes,this paper carries out the following two aspects of research:In terms of irregular text detection,aiming at the representation of irregular text components and irregular text lines,an irregular text detection method based on graph convolution is proposed.This method transforms the problem of detecting irregular scene text into the problem of generating irregular text components and irregular text lines.Firstly,the text component was generated by combining the text region with the character center points generated by the VGG16＿BN fully convolutional network,which could ensure the high recall rate of the text component and reduce the number of non-character text components.Then,the problem of grouping text lines is transformed into the problem of adjacency relation of text components,and the relationship between text components is deduced by the graph convolutional neural network.Finally,in order to evaluate the proposed algorithm,it is compared with other existing algorithms by conducting experiments on three public datasets: ICDAR2013,CTW-1500 and MSRATD500.The results show that the proposed method can well complete the detection of irregular text in scene images.In the aspect of irregular text recognition,aiming at the problem of automatic recognition of incoming water meter readings,a method based on deep learning was proposed.Firstly,the FPN network with Res Net as the backbone was used to detect the candidate regions of the water meter reading area.Then,the candidate box was combined with the character center point generated by the fully connected network based on VGG16＿BN to generate the character region,which could ensure the accuracy of segmentation and reduce the influence of noise.Then,the features of the character segmentation region were extracted by the added Roi Align layer to solve the feature extraction problem of multi-scale candidate boxes.Finally,in order to verify the effectiveness of the proposed method,the experimental demonstration was carried out through the water meter data set.The experimental results also prove that the proposed method can achieve better performance in the automatic recognition of water meter readings. |
