Design And Verification Of Natural Scene English Character Recognition Based On BNN

With the rapid development of artificial intelligence and the universal application of mobile camera equipment,natural scene images have become the most commonly used information carriers.As the most important form of image information transmission,natural scene characters have received more and more research at home and abroad.However,the existing CNN-based recognition methods are not suitable for the hardware implementation in the embedded application scenario because of the large-scale parameter.In this thesis,a binary convolutional encoder-decoder network is designed for natural scene character recognition based on the basic principle of BNN.In this thesis,firstly,the natural scene character recognition methods and the algorithm principle of CNN are analyzed,and the convolutional encoder-decoder network is designed.Then,based on the algorithm principle of BNN,the weights and feature maps in the convolutional encoder-decoder network are binarized,and the binary convolutional encoder-decoder network is further designed,then the network is trained and tested using natural scene English word grayscale images with a size of 32×128.Finally,in this thesis,the parallelism,computational partitioning and convolutional data buffering are designed in the mapping process of the binary convolutional encoder-decoder network to FPGA,and the FPGA design of the binary convolutional encoder-decoder network is completed based on Xilinx's Virtex-7 series development board VC707.On this basis,the design and test verification of the natural scene character recognition system is carried out in this thesis.The software test results show that the binary convolutional encoder-decoder network parameters is 2.14 MB,the running time on the GPU(GTX1080)is 4.59 ms,and the recognition rates on ICDAR2003 and ICDAR2013 are 92.6% and 92.1%,respectively.Compared with the convolutional encoder-decoder network,the recognition rate is slightly reduced,but the network speed is increased by 8 times and the parameters are reduced by 96%.The FPGA test results show that the recognition rates on ICDAR2003 and ICDAR2013 are 91.3% and 91.1%,respectively,at an operating frequency of 100 MHz,and the FPGA recognition speed is 33.3 frame/s.The natural scene character recognition based on BNN designed by this thesis has the characteristics of fast running speed and small memory usage.It is suitable for hardware implementation in embedded application scenarios and has certain practical application value.