| Characters are an important sign that human beings have entered the civilized age.It is of practical significance to locate and recognize characters in combination with the latest information technology.In the second decade of the 21 st century,artificial intelligence has been pushed to the forefront of public opinion.Traffic sign recognition,document recognition,and bill recognition,having been put into practical applications,make the research on character positioning and recognition of natural scenes a hot issue.Different from the positioning and recognition of document characters,characters in natural scenes are difficult to recognize due to the factors such as rich fonts,complex backgrounds,uneven lighting,foreign objects occlusion,multi-dimensional arrangement,and image distortion.Traditional methods are difficult to solve these questions.The development of deep learning technology has brought character positioning and recognition in natural scenes a new era.This thesis,combining deep learning algorithms,has studied character positioning and recognition technology in natural scenes.The main contents of the thesis are as follows:(1)In terms of character localization in natural scenes,this thesis proposes a character localization algorithm based on Cascade Mask RCNN,improving the character feature extraction network.The algorithm combines the feature pyramid network with Res Next,improves its ability to detect multi-scale features,and balances The dilemma of network complexity and accuracy.In order to adapt to the relatively large length and width of the scene character,we designed a convolution kernel with an inception structure to solve these challenges.Experiments have been conducted to verify the effectiveness of the algorithm from the perspective of recall rate,accuracy rate and comprehensive index F-score using public datasets.(2)In terms of natural scene character recognition,a character recognition algorithm based on improved CRNN is proposed.Before character recognition,the character recognition is regarded as a Seq2 Seq problem,and the space transformation based on Thin Plate Spline(TPS)is used.The network(Spatial Transformer Networks,STN)corrects irregular and perspective images,and the corrected image is sent to the convolutional neural network to extract the static features of the image.The Bi GRU network,replacing LSTM network with the complex network structure and large parameter,codes the sequence.The Bi GRU algorithm,combined with the attention mechanism,decode the sequence to obtain the recognition result of the recognized character.The ablation experiment verifies the influence of different feature extraction networks on recognition accuracy and computational overhead,also proves the optimizing effect of Bi GRU on the accuracy and complexity of the model.The ablation experiment also verifies the recognition effect of the spatial transformation network on the model on the public data set.It has an improving effect,which proves that the spatial transformation network can improve the recognition result.This thesis studies the advantages and disadvantages of existing character positioning and recognition methods on natural scene datasets,and proposes an optimized method to solve the problem of character positioning and recognition in natural scenes to a certain extent,which has reference value on scene character recognition. |
