New Product News
- This article is filed under:
- » View All
VSD Innovators Gold Award 2018 for Deep Learning Concept in FPGAs
Silicon Software GmbH Posted 06/25/2018
The concept for the implementation of pre-trained deep neural networks, specifically Convolutional Neural Networks (CNN), in FPGA processors has won the Gold Award at this year’s Vision Systems Design Innovators Awards contest. The Award was presented at The Boston Vision Show and highlights the great relevance of Deep Learning for Machine Vision and non-industrial imaging as well its use in embedded Vision.
The innovation consists of a Deep Learning Operator library within VisualApplets, a graphical development environment. By leveraging this award-winning technology, integrators and OEMs can quickly build and adapt high-speed image processing applications with big data throughput into the FPGA processor. It also simplifies the porting of CNNs to other FPGA based hardware platforms. With just a few clicks, developers can implement the Deep Learning Operator library with their pre-trained inference network using the graphical data flow models within VisualApplets. This approach dramatically reduces the development time of Deep Learning applications in FPGAs. A new CNN library operator can, for example, output the probability allocation of defect classes with the highest accuracy for an acquired image.
Accelerated Implementation of Deep Neuronal Networks (CNN) on FPGAs
Deep Learning applications and FPGAs are a match made in heaven. FPGAs offer massive parallel processing, real-time behavior with deterministic latencies and outstandingly low heat generation, which is ideally suited for Deep Learning in industrial and embedded requirements. Acquired images are processed in real-time by the frame grabber’s FPGA where big bandwidth capacity can generate high performance and detection rates of over 99%. This processing capability allows high resolution images to be classified in real-time. This technology also benefits developers with the long product lifecycle of FPGAs as well as their CNN Intellectual Property being protected within the FPGA hardware.
A “CNN-optimized” frame grabber with a larger and more powerful FPGA is currently in development. It will permit even more complex Deep Learning applications with even higher data throughput in real-time applications.