Aiming at the problem that limited on-chip resources of FPGA restrict the implementation of complex neural networks

by using dynamic partial reconstruction technology

two hardware and software schemes are designed. Compared with pure software scheme

the network execution speed is increased more than three times faster. The self-made license plate character dataset achieves 99. 45 % training precision under the Caffe network framework. Finally

based on PYNQ—Z1 development board

the design of fast license plate recognition system is completed.