We all like to talk about the super high-end of machine learning with computer vision algorithms running on a turbo-charged, 10 tera-operations-per-second accelerator, but the reality, especially for our embedded industry, is that the majority of applications need a processing engine suitable enough to get the job done and no more. This is our motivation for offering scalable machine learning devices from MCUs (such as the Arm® Cortex®-M7-based i.MX RT1050) to application processors (such as the i.MX 8QuadMax and Layerscape® LS1046) – and finally you’re able to see this range of performance in action with no less than 12 machine learning demos at the Arm TechCon in the NXP booth (details below).
For example, stop by the booth and see a wide range of solutions representing low cost, low power, secure, and high performance face recognition. How about face recognition solutions starting at $2 USD? Our design starts with an NXP i.MX RT1020, a low-cost device sporting an Arm® Cortex®-M7 core. NXP developed its own face recognition algorithms, and the ability to train for new faces directly on the RT1020 platform. The outcome is face detection and recognition in slightly more than 200msec with accuracy up to 95% – starting at $2 USD. Higher performance face recognition examples will also be on display using devices such as i.MX 7ULP (high-performance and ultra-low-power), i.MX 8M Nano (real-time face detection using Haar Cascades to give an efficient result of classifiers), and i.MX 8M Mini (doing secure identification with anti-spoofing), and the i.MX 8M Quad-based Google® Coral Board with the Google TPU (for super-fast facial recognition in a sea of people).
Moving on to image classification, the NXP booth will host an application using the i.MX RT1060 and the eIQ™ machine learning software development environment. This example performs classification with a TensorFlow Lite model trained to recognize different types of flowers (sunflower, tulip, rose, dandelion, and daisy). Specifically, we’re running a MobileNet model and doing inferencing at the rate of 3 frames per second – on an MCU! This demonstration also shows the flexibility of eIQ, providing support for a variety of inference mechanisms (e.g. TensorFlow Lite, CMSIS-NN, Glow) and other types of machine learning models besides image classification (e.g. audio or anomaly detection).
Other Cool NXP Things at Arm TechCon