The ARM architecture is one of the most popular and successful CPU architecture today. The relatively low power consumption of ARM CPUs makes it an architecture of choice for edge computing on embedded devices. This is also true in machine vision, and many ARM-based smart cameras and computing devices such as the Nvidia Jetson are available. In particular, Euresys is working with several smart camera manufacturers to allow the most efficient implementation of the Open eVision libraries on their platform. The new version 22.04 runs on ARMv8-A compatible processors with at least 512MB RAM and 512MB of storage, and running 64-bit Linux. Only C++ development is available. Direct compilation on the embedded device requires at least 4GB RAM. Cross compilation on a Linux x86 PC is possible. Several sample programs, using the console or based on Qt, are available. Additionally, it is possible to build, learn, optimize and save models (such as deep learning models) using Open eVision Studio on Windows, then load them from Linux ARM applications. All 18 libraries of Open eVision are available. The 3D libraries include laser line extraction and calibration functions, point cloud processing and management functions, and 3D object extraction, alignment and inspection functions. The Deep Learning libraries are CNN-based inspection libraries for image classification and segmentation. The general-purpose libraries cover applications such as image filtering and enhancement. The matching and measurement tools contain blob analysis, pattern matching, alignment and sub-pixel measurement functions. The text and code reading libraries include functions for OCR and 1D/2D barcode reading and grading.
Beside ARM-based smart cameras, Open eVision 22.04 has also been validated on Raspberry Pi and Nvidia Jetson boards. The various Nvidia Jetson boards are complete System on Module (SOM) that include multiple-core GPUs. This hardware makes the Jetson boards suitable for training and inference phases in deep learning applications. GPU-accelerated training and inference are not available yet in this version of Open eVision’s deep learning libraries (EasyClassify, EasySegment and EasyLocate), but they will be available in the next release (using NVidia CUDA). However, CPU-based inference and training are already available.