Deep learning computer vision algorithms ported to processor IP

Jon Lawson

Synopsys has teamed up with Morpho to port computational photography software to the DesignWare EV6x Vision Processor.

Morpho's Scene Classifier image classification technology uses deep learning algorithms to analyse visual input and automatically apply tags for classification, searchability and organisation.

Morpho is optimising their software to take advantage of the EV6x Vision Processors' scalable hardware architecture, which includes a 32bit scalar core with up to four 512-bit vector DSPs and a fully programmable convolutional neural network (CNN) engine.

The combined hardware-software solution enables designers to accelerate image classification and automated tagging tasks in their mobile and surveillance systems-on-chips (SoCs) while consuming significantly less power and memory resources, allowing the cores to be more effective in system-on-chip designs.

"We see growing demand for image processing software that takes advantage of deep learning networks to reduce computational resource requirements, particularly for battery-powered mobile devices," said Masayuki Urushiyama, executive vice president at Morpho. "Optimizing our deep learning and image processing software for the DesignWare EV6x Vision Processors enables designers to implement high-quality image recognition and classification solutions that increase the vision processing capabilities of their SoCs and consume much less energy than traditional GPU approaches."

Morpho's Scene Classifier uses deep learning to ‘recognise’ essential identifying features for automated, real-time image tagging. Morpho's software algorithms include high-precision scene recognition technology, motion detection, 360 virtual reality stitching technology and other image processing technology.

The DesignWare EV6x Vision Processor IP is supported by a comprehensive software development environment including the ARC MetaWare EV Toolkit to address a wide range of automotive, industrial and consumer applications.

The EV61, EV62 and EV64 cores are capable of up to 4.5 TeraMACs/sec when implemented in 16nm processes under typical conditions and support multiple camera input with resolutions up to 4K.

The EV6x supports any CNN, including popular networks such as AlexNet, VGG16, GoogLeNet, Yolo, Faster R-CNN, SqueezeNet, and ResNet.

"The emergence of deep learning for image classification, detection and recognition enables a new level of image processing efficiency in SoC designs," said John Koeter, vice president of marketing for IP at Synopsys. "By collaborating with Morpho to optimise their software for our EV6x Vision Processors, we are providing designers with a hardware-software solution that significantly improves the accuracy, performance and power consumption of image processing in power-sensitive applications."

The Morpho Scene Classifier software optimised for EV6x Vision Processors is planned to be available in Q4 2017.