A Research On Lao Language Text Recognition With Multi-feature Fusion

Lao is the script of Lao People's Democratic Republic,As an one important allied country with "the Belt and Road Initiative",The research of Lao is of great significance,Due to the small population and backward digitalization in Laos,it is difficult to directly obtain Lao text corpus from the Internet.However,there are a large number of Lao text images.Therefore,how to accurately recognize Lao characters from the existing text images to expand Lao text corpus has become one of the research focuses of Lao natural language processing.OCR can effectively extract the characters contained in the text image,However,there are few and far between researches on Lao character recognition at present.Therefore,this paper refers to the relevant text recognition research,and proposes a Laos text recognition method based on multi feature fusion to solve the difficulties in the current Lao language research work.It mainly includes the following three parts:(1)Construct ResNet-BiLSTM-CTC end-to-end Lao character recognition model.Because Lao sentences are composed of characters,and most of the characters are composed of the upper,middle and lower structures,the model fusion tone features and lower consonant features to assist the model to accurately recognize Lao characters.Firstly,the deep convolution network is used to extract more abundant character structure information,and the Bounding-Box of single character is corrected by border regression,.Then the BiLSTM network is used to characterize more advanced semantic information.Finally,the prediction results are transcribed by CTC to obtain recognition sequences.Compared with Deep BLSTM,DenseNet,Text-CNN and other models,the validity of ResNet-BiLSTM-CTC model is confirmed,and the accuracy rate reaches 80.45%.(2)Based on the ResNet-BiLSTM-CTC model,In view of the character breakage and adhesion caused by the quality of text image,The text puts forward the "Lao character writing level",and construct a Multi-task Learning Model Combining character recognition Loss and character writing grade assistant loss,Firstly,ResNet network is used as the shared feature extraction layer.One side is used to segment Lao single characters,and the other side is used to fit the writing level of Lao characters.Then after the two vectors are spliced with the input text image vector,the Lao character sequence is recognized by BiLSTM,and finally CTC transcription is carried out to obtain the recognition result.After experiments and comparison with other mainstream methods,it is proved that the ResNet-BiLSTM-CTC model is effective after fusing Lao character writing grade features,and its accuracy rate reaches 86.94%,which is 6.49%higher than the previous work accuracy rate.(3)In order to obtain more accurate Lao character recognition results,error analysis was carried out on the previous recognition results,and it was found that the model could not distinguish Lao similar characters well.Therefore,the problem of distinguishing similar characters was modeled.Spatial and channel attention mechanisms are introduced in different stages of the model to focus on the local morphological information of characters and the context information of sequences.And Unicode table is constructed to solve the problem that similar characters in Lao are easy to confuse,which helps to distinguish the visual similarity between character pairs.In order to prove its effectiveness,the effectiveness of the model is tested by the recognition of Lao printed,handwritten and scene text images.The experimental results show that the accuracy rate reaches 90.45%.