A Flexible Machine Vision AI System for Edge-Oriented Deep Learning Accelerators

In recent 10 years, deep learning has successfully shown its effectiveness in various computer vision fields such as autonomous vehicles, robotics, and AI surveillance. Numerous machine vision AI systems have been accordingly developed to run those deep learning algorithms. However, existing PC-based machine vision AI systems have the disadvantage of having to modify the entire system although just a small change is required. They are also considerably high cost/power consuming systems for edge device purpose which is targeted in this study. In addition, a number of FPGA-based machine vision AI systems are not suitable for multiple applications as they are dedicatedly designed. In this work, in order to overcome those disadvantages, we have developed a flexible FPGA-based machine vision AI system where various type of accelerator can be implemented. An accelerator integrated with a post processing unit(PPU) that runs an object detection was successfully demonstrated on the system satisfying the required input image resolution/format. The proposed system can also provide multiple resolution/format of input images which is proved via Xilinx integrated logic analyzer(ILA) debugging. We therefore ensure that this system can further support more accelerators which can be deployed in diverse applications.