XM4 DSP has been enriched with CEVA Deep Neural Network (CDNN) Software Framework. Some explanation could be useful before jumping into CDNN. The “Deep” of CDNN comes from “Deep Learning”, a family of neural network methods using high number of layers, so a deep network. The most popular deep learning neural network method is the “Convolutional Neural Network”. Why is it popular? Because CNN focus on feature representations, required to support applications like object recognition, driver assistance (ADAS) or augmented reality, to name a few emerging applications becoming very popular, generating developments in various segments, from automotive to consumer. CNN offer two major benefits, justifying this infatuation. At first, CNN provides best recognition quality when compared with alternative recognition algorithms. The second benefit is linked with the artificial intelligence nature of the algorithm: the designer will implement it once and be able to use it many times without code change, through re-training. Such a benefit could greatly accelerate machine learning deployment for embedded systems, as, by definition, you want such systems to run as long as possible without intervention.
CEVA has run a partnership with Phi Algorithm Solutions and optimized the CNN-based Universal Object Detection (UOD) algorithm from Phi, and ported it to CEVA-XM4 via CDNN. The first “N” of CNN is for Neural, indicating that researchers strive to mimic the human brain in computers. Such work was limited mainly by computing horsepower, power constraints and algorithmic quality, but the technology progresses allow to bring neural network in the embedded world. Harnessing the computing power of the CEVA-XM4 imaging & vision DSP, the partners have created the lowest power and memory bandwidth deep learning solution providing real-time, efficient object recognition and vision analytics.
The concept of pre-trained networks is brilliant: the designer receives network model & weights as design inputs from offline training (pre-trained) and these are automatically converted into a real-time network model, via CEVA Network generator. He can utilize this real-time network model in CNN application on CEVA XM4 DSP. This usage flow is described below (Caffe is a popular open source software framework, used to build, train, activate neural networks).
If you look at the usage flow, CEVA main contribution is the Network Generator, allowing merging two distinct know-how. The 100% software based science using floating-point algorithms on the left side has to be converted into fixed point, power aware and hardware compatible customized network to be implemented into an embedded DSP… keeping high recognition accuracy. CEVA claims less than 1% degradation in accuracy compared to the original network, which means that in less than 1% of cases, the pictured Labrador retriever could be confused with a Beagle.
CNN-based Universal Object Detector algorithm (from Phi Algorithm Solution) is now available for application developers and OEMs to run a variety of applications including pedestrian detection and face detection for security, ADAS and other embedded devices based around low-power camera-enabled systems.
Taking the example of pedestrian detection the real-time detection application utilizing CDNN and optimized for CEVA-XM4 DSP exhibit less than 30 mW for 1080p, 30 fps and provide 15x average memory bandwidth reduction compared to typical neural network implementations. This makes it the lowest power deep learning solution for embedded systems: 30x lower power and 3x faster processing when compared to leading GPU-based systems.
According with Eran Briman, vice president of marketing at CEVA, “Our new Deep Neural Network framework for the CEVA-XM4 is the first of its kind in the embedded industry, providing a significant step forward for developers looking to implement viable deep learning algorithms within power-constrained embedded systems.” The CEVA-XM4 imaging & vision DSP together with the CDNN framework paves the way to advances in artificial intelligence devices in the coming years using deep learning techniques.Share this post via: