Research On Image Text Classification And Translation Technology Based On Neural Networks

In the field of processing image information,image text classification and translation is an important research direction and widely applied in industries and our daily lives.At present,based on the characteristics of text information itself,different processing tasks are faced in information processing.Each problem has its own feature information and requires different processing schemes.Therefore,this article takes person name classification,news topic classification,and image text translation as examples to design corresponding models to optimize information processing and provide corresponding services for industrial needs.This article adopts neural network technology,focusing on the topic of image text classification and translation,with text information as the research object,and proposes corresponding solutions on how to better process image text information.Aiming at the research and application of image text information processing,in order to facilitate the subsequent processing of image text information,we need to solve the problem of image text recognition,so we designed an image information recognition model based on Convolutional neural network,which consists of feature extraction layer,cycle layer,transcription layer,etc.In the optimization,CTC Loss function is used to replace the previous optimized Loss function.The experiment shows that the image text can be recognized.Next,in order to how to better process text information,we designed a long-term and short-term Recurrent neural network(LSTRNN)to better process name information.LSTRNN is composed of traditional Recurrent neural network and long-term and short-term memory network.The process is to use continuous Bag-of-words model to obtain word vectors,LSTRNN layer to extract name features,and full connection layer to integrate information to achieve nationality classification.A news text processing model based on Bi GRU is designed.The process is to capture the association between high-dimensional spatial words through Word embedding,obtain the word vector representation,and Bi GRU captures the semantic information between consecutive sequences to obtain the comprehensive feature representation of news text,which is classified through the full connection layer.In terms of optimization,tag smoothing,adjusting the Learning rate adjustment method Step LR and Logsoftmax,are introduced.A Transformer based translation model was designed to handle image text translation.Adapt its encoder and decoder,decode the output and input decoder obtained by the encoder,and generate the sequence to combine with the Greedy algorithm to achieve translation.Improved translation accuracy by optimizing the internal sub layer output mode of the original Transformer structure and introducing label smoothing.The experiment shows that the designed LSTRNN model can improve the accuracy of person name nationality classification,with an acc of 98.58%,acc@topk3 99%;The accuracy of the news subject classification model can reach 91.1%;The image text translation model can effectively extract the correlation features between languages and translate them,achieving relatively good results,with a BLEU of 38.5 on the self-collected English French dataset.The above scheme results all verify the correctness of the research in the paper and meet the purpose of better processing image text information that meets the needs of the industry.