Design And Research Of Certificate Text Information Detection System Based On FPGA OpenCL

Document text information detection belongs to the research scope of text image detection,there are many achievements in this field at present.Although the algorithm structure of traditional image processing method is relatively simple,but it's generalization ability sometime is poor and in some complex scenes,it can not detect accurately.For deep learning algorithm,because of its complex network structure and huge parameter calculation,it has a high dependence on the performance of the system hardware.FPGA devices have the characteristics of small size and high energy consumption ratio,which attract more and more attention in the field of parallel computing and deep learning.However,the traditional method of developing FPGA acceleration program has high threshold,complex process and long cycle,which can not adapt to the change of scene and requirement quickly.OpenCL is a kind of parallel computing framework first developed by Apple company.It is convenient and fast to implement parallel algorithm acceleration system by using advanced programming language.At the same time,because of its cross platform and multi device characteristics,the parallel acceleration system developed in the OpenCL specification framework can be transplanted to different hardware devices quickly to achieve a high degree of software compatibility,and according to different demand scenarios,we can choose different hardware supported by OpenCL to realize the deployment of the system rapidly.This paper designs and implements an image text information detection system based on FPGA platform,which conforms to OpenCL programming specification.The system uses CTPN text detection network as the core algorithm model,and in the process of system design,this paper deeply studied the algorithm structure of CTPN network and the principle of OpenCL parallel acceleration.Under the framework of OpenCL parallel programming,this paper presents a simple,efficient and fast parallel vector accumulation algorithm based on the idea of merging,which greatly improves the efficiency of data accumulation.At the same time,the parallel vgg16 convolutional neural network,the parallel bilstm bi-directional recurrent neural network and the parallel RPN region proposal network algorithm for CTPN network are proposed.Compared with the serial algorithm,it can significantly improve the operation speed,and the performance of a single three-dimensional convolution operation is up to 142 times higher,the performance of parallel operation in Bi LSTM network is improved by more than 10 times,at the same time,by integrating the parallel algorithm,the overall running speed of the system is increased by 88.6 times,with large amount of test data.For the detection effect of the system,this paper uses five different types of text document images,such as ID card,ticket,bank card,graduation certificate and business license,to test and verifies the effect of text detection and generalization ability of the system in different scenarios...