Hardware Acceleration of Computer Vision and Deep Learning Algorithms on the Edge using OpenCL

Abstract

Machine vision using CNN is a key application in Industrial automation environment, enabling real time as well as offline analytics. A lot of processing is required in real time, and in high speed environment variable latency of data transfer makes a cloud solution unreliable. There is a need for application specific hardware acceleration to process CNNs and traditional computer vision algorithms. Cost and time-to-market are critical factors in the fast moving Industrial automation segment which makes RTL based custom hardware accelerators infeasible. This work proposes a low-cost, scalable, compute-at-the-edge solution using FPGA and OpenCL. The paper proposes a methodology that can be used to accelerate traditional as well as machine learning based computer vision algorithms.