Design And Development Of Mobile OCR Model Based On Network Architecture Search

With the development of intelligent mobile technology,the application of OCR technology on mobile devices has brought great convenience to people's daily life learning and meets the growing learning needs of people.However,the existing mobile terminal OCR technology relies excessively on network communication,that is,it needs the support of the cloud terminal service,and is usually not available offline.Although the mobile OCR recognition technology is flexible,it is not possible to achieve the application-level recognition effect by embedding the existing OCR model directly in the mobile terminal.Designing a lightweight mobile terminal OCR model search algorithm that can take into account the recognition accuracy,model size and time-consuming,has high practical significance and practical value.Regarding the issue above,the paper takes the design and development of mobile terminal OCR model as the main research and development content.The basic idea is that the algorithm should start from the current high-performance OCR technology and solve the technical difficulties of embedding the high-performance OCR model into the mobile terminal.Through the network structure search,the OCR model suitable for the mobile terminal is searched under the constraint of the given objective function.Main work of the paper:Firstly,study the research status of OCR technology at home and abroad,analyze relevant research results,and establish the research content of this paper;Based on the research of OCR algorithm framework,the technical route of OCR model design and development is determined.Secondly,the best deep learning technology in the field of computer vision is used as the infrastructure of OCR model design.Through the analysis of the existing deep learning-based OCR model,and through experiments,the deep learning that this paper expects to embed into mobile terminal is determined.After that,based on the neural network lightweight strategy,the selected depth model is lightly improved,and the storage space occupied by the model and the time consumption of the model are reduced as much as possible without loss of precision.Then,using the network structure search strategy,search and confirm the lightweight model structure parameters that meet the constraints.Finally,the validity of the search structure is verified based on the public data set collected by the real environment.The OCR model of the mobile terminal searched in this paper was tested on the actual mobile phone mobile terminal.The test results show that the OCR model searched by the search strategy takes into account the model size,time consumption and recognition accuracy,and achieves the expected effect.It basically meets the needs of daily text recognition tasks and reflects certain practical application value.