Design of Handwritten Mongolian Recognition System Based on FPGA

Recently, the AlexNet network model has achieved excellent performance in image processing. However, this model is difficult to apply in the embedded field, because there are huge data parameters and complicated calculations. Therefore, we modified the AlexNet network to reduce the number of parameters to 18,532 and used the standard dataset MNIST to test the network, and the accuracy reached 97.57%. We have built a map library of Ordinal Numbers in Mongolian. By collecting script Mongolian word order numbers, image preprocessing, and classification, we completed the establishment of the Mongolian ordinal image library. The database test results show that the accuracy rate is 96.66%. Finally, we deploy the improved AlexNet network to the FPGA to the recognition system. System tests showed a correct rate of 95.62 percent, an identification speed 11.07 times higher than a PC system, and a system power consumption of 0.405W.