| Scene text recognition refers to the process of extracting text information from natural scene images.Recognition of text in natural scenes has important theoretical research significance and application value.At present,automatic driving,image retrieval,real-time translation and other application scenarios need the support of scene text recognition technology.However,there are still many challenges in the current text recognition technology,such as the recognition accuracy is not good when the text image is occluded by watermark,the academic circles pay insufficient attention to Chinese text recognition,and the lack of Chinese and English recognition models that balance speed and accuracy.To address these issues,the main contributions of this thesis are as follows:1.The problem that the recognition accuracy is poor when the text is occluded by the visible watermark has been solved.On the one hand,a text watermarked image dataset is constructed,which simulates the situation of watermarks that may appear in the real world.The training set is used to train the text image watermark removal model,and the test set is used to evaluate the effect of the text image watermark removal model.On the other hand,a new text image watermark removal model is proposed.The new model is improved based on the general watermark removal model,and the main innovations are as follows:Firstly,a sequence modeling module is added to the generative network,and the sequence modeling module utilizes the contextual relationship to better recover the text occluded by the watermark.Secondly,the text recognition model is used as a perceptual loss to train the generative network,so that the watermark removal task is organically combined with the downstream recognition task to improve the recognition accuracy.Finally,the L1 loss is replaced by the masked L1 loss,which forces the generation network to focus more on the watermark area,thus improving the watermark removal effect.2.We propose a text recognition model that balances speed and accuracy.The model is improved from a Connectionist Temporal Classification(CTC)based recognition model.The main innovations are as follows:Firstly,two problems existing in the CTC based method are solved.We eliminate the feature degradation problem caused by multi-path in the CTC decoding process by introducing a feature learning loss.By adding an adjacency modeling layer,we eliminate prediction errors caused by misalignment of character and feature sequences during CTC decoding process.Secondly,in order to solve two problems of recurrent neural networks(RNN),that is,the slow running speed caused by the difficulty of parallelization,and the accumulation of errors caused by sequential prediction,self-attention mechanism is used to replace RNN in the sequence modeling layer,which improves the running speed and recognition accuracy.Finally,by introducing the Focal-CTC loss,the problem caused by the long-tail effect of the distribution between Chinese characters is solved,and the recognition accuracy of rare characters has been significantly improved.3.Two text recognition systems are designed and implemented,which verifies the validity of the model proposed in this thesis and can be applied in actual needs.One is the ID card element extraction system under the interference of watermark,the system can extract the personal information in the copy of ID card under the interference of watermark.The other is a marking system for teaching assistance,which can realize the automatic identification of class,student number,and answering results of objective questions,and store the identification results in a structured manner to facilitate subsequent analysis and use.We have also developed a user interface for the marking system for teaching assistance,which is userfriendly and supports visualization and correction of recognition results.The both system systems have high accuracy and initially meet the design objectives and daily use requirements. |
