Version 17.12 of HALCON ships with two pretrained networks, one of which is a "compact" network that is optimized for speed and suitable for use on embedded boards. This software was successfully tested on NVIDIA Jetson TX2 boards based on 64-bit ARM processors. The deep learning inference, i.e., applying the trained convolutional neural network 9CNN) almost reached the speed of a conventional laptop GPU (approx. 5 milliseconds). This, according to MVTec, is an unusually-high performance for an embedded device compared to a standard PC.
As a result, users can now deploy deep learning techniques on the NVIDIA Jetson TX2 embedded board. MVTec will provide interested customers with a software version for this architecture on request. In addition to deep learning, HALCON’s standard machine vision library is available on embedded devices. Applications can be developed on a standard PC and transferred to an embedded device via HDevEngine. Furthermore, notes MVTec, users can utilize more powerful GPUs, available for the PC to train their CNN and execute the inference on the embedded system.
"We have provided successful technological proof that allows us to offer advanced deep learning functions in the embedded vision segment. This will greatly benefit users. They can now utilize the extensive new HALCON 17.12 features on standard devices with NVIDIA Pascal architecture – at an extraordinary high speed for embedded technologies," explained Christoph Wagner, MVTec's Embedded Vision Product Manager.
Dr. Olaf Munkelt, Managing Director of MVTec Software, also commented: "The rapidly growing market for embedded systems requires corresponding high-performing technologies. At the same time, AI-based methods such as deep learning and CNNs, are becoming more and more important in highly automated industrial processes. We are specifically addressing these two market requirements by combining HALCON 17.12 with the NVIDIA Pascal architecture."
View more information on HALCON.