Research On Low-quality Character Recognition Method Based On GAN

Optical Character Recognition(OCR),as an important branch in the field of computer vision,has a wide range of application space and research value in both natural scenes and specific scenes.Traditional character recognition methods have achieved Great success in character recognition tasks with simple backgrounds such as documents,but they are inevitably stretched for character recognition in more complex specific scenes.In recent years,with the continuous deepening of research on convolutional neural networks,because it can easily extract the deep feature relationship between data,it greatly improves the processing efficiency of image data,and is widely used in character recognition tasks.However,through research,it has been found that the existing deep learning character recognition methods have insufficient recognition results when faced with "low-quality" character data with incomplete morphology due to various reasons.Because the deep learning method requires a large amount of data to establish a reliable mapping relationship between the recognition target and the label,but in actual situations,the frequency of "low-quality" character data is low,and it is difficult to add a large amount of data to subsequent training.Improve the recognition accuracy of the neural network.For such situations,there is currently no better solution.To this end,this article proposes a new solution,that is,firstly generate a large amount of "low-quality" character data by generating a confrontation network,and obtain a mixed data set of it and real character data;secondly,by designing a lighter with stronger Generalization ability The magnitude convolutional neural network trains and recognizes the data.The main work is as follows:(1)Design a generative countermeasure network to generate "low quality" data.First,the model is trained using normal character data and semantic sketch data.After the model is trained,low-quality character data is generated by drawing semantic sketches of "low-quality" data.Two kinds of industrial characters were used as examples to generate low-quality data.At the same time,a comparative experiment using the traditional generative countermeasure network proved that the designed network can not only guide the data generation of "low-quality" data artificially and efficiently,but also make the whole process more interpretable.(2)Design a lightweight network suitable for character recognition tasks.Since "low-quality" data is a problem derived from actual application scenarios,it makes sense to be able to deploy to actual scenarios.Based on the calculation amount and memory access cost and other factors,this paper selects the detection head,backbone network,feature optimizer,loss function and other structures suitable for the task when designing the lightweight network,and improves the part of the excessive calculation.,To ensure that the entire network has extremely high deployment friendliness.The improvement in the detection head loss function improves the quality of the network character detection boundary,and lays an important foundation for the performance of the network in character recognition tasks.Finally,the normal data and the generated "low-quality" data are trained on the lightweight network,and comparative experiments are designed and the recognition effect is tested.The character recognition method proposed in this paper is not only suitable for industrial character environment,but industrial character data is used as an example to show the experimental effect.In fact,"low-quality" data will appear in various fields in the application environment,which has become a major difficulty in the recognition of various fields(for example,characters that are missing due to lighting or occlusion in natural scenes).As long as the above problems are adjusted to a certain extent,the accuracy of the model's recognition of low-quality data can be improved through the method of this article.