Lightweight convolutional neural network of YOLO v3- Tiny algorithm on FPGA for target detection

With the rapid development of artificial intelligence, the convolution of the neural network has been deploying across a range of industries, and has certain applications in national defense security, transportation monitoring and medical research. However, due to the constraints of speed and the power consumption, convolutional neural networks are still limited to a great extent in the implementation of the edge computing and mobile devices, etc. For this reason, we design a lightweight convolutional neural network based on FPGA. In this paper, we use the YOLOv3- Tiny algorithm, which is fast in execution, small in computation and small in size, is suitable for deployments on embedded devices such as FPGA. This paper uses 16-bit fixed-point quantization, special data storage and hardware circuit for convolutional computation written in verilog - hardware description language, consumes 512 DSPs, consumes 37.037 ms to recognize a single image frame, and consumes 9.611W. The system basically achieves the design goal of target detection.