New-Tech Europe Magazine | July 2017

Stack based solutions for image processing at the Edge and Cloud

Nick Ni & Adam Taylor, XILINX

One of the beauties of image processing is its wide range of end applications from autonomous drones and cars, to medical and scientific imaging. This diverse range of applications brings with it a diverse range of requirements and solutions, from embedding the intelligence at the edge, to leveraging the power of the cloud. In either case, for their image processing system, users will face several challenges. This article looks at what these challenges are and how they can be addressed using an acceleration stack based approach. Challenges Applications processing at both the Edge andCloud initially face a common problem, which is the implementation of the image processing algorithm such that it meets its overall system requirements. There will, in most

these classifiers within its application. It can be seen then, that both implementations require the capability to work with modern frameworks such as OpenCV, OpenVX, Caffe and FFmpeg to achieve their image processing requirement. But what about other requirements which dominate in these different implementations, these requirements must also be considered and addressed. Processing within the Edge brings with it not only the need for real time processing and decision making but also its applications are often autonomous which brings other challenges. Autonomous operation brings a need for both a secure system and secure communications channels (when available) back to its operations centre. Autonomous systems are also often battery

cases, be a difference as to what that driving overall system requirement is. For an edge based implementation it may be the latency of the algorithm as the system may be required to make decisions based on information contained within. While a cloud based image processing solution may be driven by the requirement for exceptional accuracy as scientific or medical decisions may be based upon this. Both implementations will also heavily rely upon deep machine learning and artificial intelligence, although in different manners. Edge based processing will use the classifiers generated by deep machine learning within the cloud to implement its object detection algorithms. While cloud based solution will use deep machine learning and Neural Networks to both generate the classifiers and then use

26 l New-Tech Magazine Europe