CNN for object recognition implementation on FPGA using PYNQ framework

Object recognition is one of the most researched and commercialized applications of Deep Learning (DL) where Convolutional Neural Networks (CNNs) are especially accurate. The deployment of these models on embedded systems require low latency and high performance even with limited resources and energy budgets. Embedded systems with Zynq Systems on Chips (SoCs) are attractive platforms for CNNs. In this paper, we use PYNQ framework, that supports a Python-based hardware/software codesign environment to perform CNN inference for object recognition on Xilinx FPGA. We design the CNN model and train it on a CPU platform and we implement it On ZedBoard FPGA. By using only, a single ARM processor core on FPGA, we achieve 100ms latency and up to 10 image recognitions per second on the CIFAR-10 dataset with 79.90% accuracy. This model performance can be highly improved by exploring the hardware resources of the FPGA chip.