Implementation of post processing hardware for real-time object detection in CNN acceleration system

Abstract

In neural network based real-time object detection and recognition systems, the scaling of layer processing acceleration hardware can be one way to achieve realtime goal. However, this approach places burden on hardware resources, and consequently requires bulky hardware on the system. In this paper, we propose a post processing hardware that can improve the operation speed of an application while using less hardware resources. The proposed hardware implements the most time-consuming post processing part in our previous study. The proposed hardware achieves about 93.8% better than the conventional method in Xilinx Zynq Ultrascale + MPSoC ZCU106 processing system (PS).