The hardware acceleration system was deployed on Xilinx FPGA chip xczu9eg, with a clock frequency of 180 MHz. The GPU used was NVIDIA GeForce GTX 1080 Ti. The execution times of ShuffleNetV2 on FPGA and GPU are shown in the table Ⅱ, representing the average time for classifying 1250 images. The speed on FPGA is 1.12 times faster than that on GPU, with a decrease in recognition accuracy of only 0.66%. Since it takes 146M FLOPS to perform the whole network, we can calculate the GPU’s energy efficiency as 0.44. Our work has achieved an energy efficiency 30.57 times higher than that of the GPU.