Software and Hardware Cooperative Acceleration Technology for CNN

Abstract

To meet requirements of building intelligent avionics systems, and improve the intelligent combat level of manned/unmanned aerial vehicles, the software and hardware cooperative acceleration technology for CNN is designed and implemented to solve complex problems such as target recognition, auxiliary decision-making, and autonomous planning. Aiming at solving the conflicts between the huge amount of parameters and the limited storage resources for embedded environment, the neural network model is optimized with model structure optimization and quantization of parameters. Aiming at solving the conflicts between complex floating-point operations and the shortage of computing resources, the convolution accelerating operator and the pooling accelerating operator are designed based on Verilog HDL. The pipeline and full parallel method are used to achieve the purpose of acceleration. Through the synergy of software optimization and hardware accelerated, the inference process of convolutional neural network is accelerated. Two popular CNN networks, that are YOLOv3 and YOLOv3-Tiny, are used as examples to accelerate and verify on the Xilinx ZCU102 FPGA development board. The results show that compared with the original models, the parameters of the accelerated models can be compressed about 3/4. The inference speed of YOLOv3 is increased by nearly 65 times, and that of YOLOv3-Tiny is increased by about 23 times.

Keywords

• |intelligent computing|hardware acceleration|target detection|model compression|fpga