AI-Box - Smart Visual Inspection
deevio automates visual inspections in production. Our core product is an artificial intelligence-based machine vision software with which we help manufacturing companies to automate their still manually performed quality inspections and to significantly improve their existing machine vision systems.
​
We supply our customers with a camera system tailored to the individual application, which can then be integrated into the respective machine environment via a PLC connection.
​
The experience we have gained over the past years by implementing inspection cases, has helped us to develop hardware and software to meet all the requirements of an AI-driven inspection workflow.
GigE Vision and GeniCam
Supporting cameras through industry standards means you can choose from a wide range of options of grayscale, color, or even 3D cameras to suit your use case.
Power over Ethernet (PoE)
With PoE, cameras can directly connect data and power with a single cable, supporting up to 100m in length.
Real-time GPIO’s
A dedicated FPGA enables real-time control of GPIOs for optimal setup of triggers for light and camera as well as other external peripherals.
1. Acquire & Label Images
Any visual inspection starts with taking images through cameras. With the AI-Box, cameras and lights can be directly connected without the need for intermediate switches or controllers. Using High-Dynamic-Range (HDR) photography, even highly reflective surfaces can be acquired with sufficient local contrast.
​
The AI-Box also has labelling software that allows you to draw a detailled label on the actual defect in the image. This way, you teach our deep learning software what makes a defect and what not.
2. Inspect Images
The core of the AI-Box runs a deep learning algorithm that has now learned how to distinguish images from “OK” and “not OK” products. Based on the images and labels from the first step, deevio trains a deep learning model that is tailored to your use case.
​
Unlike traditional rule-based inspection, our AI is trained on those labeled images that you provided.
This enables a new workflow where the bottleneck is no longer the vision engineer coding the inspection, but where the AI itself learns how to inspect.
Binary and multi-class classification
for each input image, the AI algorithm returns a single OK/NOK answer. Or you extend the training data by including different error classes, and our algorithm learns to distinguish multiple types of defects
Image
Segmentation
for each input image, the AI algorithm can show exactly where the defect is in the image.
Anomaly
detection
in cases where the defect covers a large part of the image, our algorithm recognizes when an image contains an anomaly not seen before. This algorithm can be trained utilizing only images of good parts and does not require the customer to provide images of defects.
3. Manage Your Inspection Case
AI-based solutions require a dedicated workflow, this is supported by the AI-Box:
1.
Create a new product, add classes or defect types you want to identify, and start acquiring and annotating images.
2.
Review images, remove wrong labels or adapt annotations.
3.
Let deevio create a deep neural network-based model that is tailored to your new product and upload it to AI-Box
4.
Change your inspection to the new model and start inspecting on the production line!​
Create, review, adapt and export data sets that are ready to train an AI model.
​Annotate images with use case-specific labels. Draw on top of images with our labeling tool to mark the exact location of a defect.
​Data export: No lock-in, you can easily export images and annotation data from AI-Box
Specs
The AI-Box application is optimized to run on the AI-Box, a small and efficient machine vision PC that is built to our design.
​
In addition, the software stack runs on a wide range of NVIDIA-GPU-enabled PCs to serve any requirements for cycle time, image resolution, and complexity of the AI-models.
Low energy
WIFI, 4G
2xPoE/1x Gigabit Ethernet
Digital IO
24V DC
General
System on a Chip (SoC)
Nvidia Tegra X2, 64 Bit
CPU
4-core ARM Cortex-A57, 2GHz
2-core Denver2, 2GHz
GPU
256 CUDA cores (Pascal), 1.12 GHz
RAM
8 GB DDR4
Storage
1 x M.2 SSD (256 GB)
1 x 32 GB eMMC
1 x SDXC slot
Interfaces
Camera Interface
2 x GigE Vision with Trigger-and-Power-over-Ethernet
Camera Trigger
2 x Trigger-Over-Ethernet
Digital I/O
8 x Input & 8 x Output, opto-isolated, 24 V
Encoder Interface RS-422
3 x input for one encoder
LAN
1 x 1000 MBit/s
Wireless LAN
1 x IEEE 802.11a/b/g/n/ac with external antenna socket
LTE (4G)
1 x internal socket for SIM, with external antenna socket
USB
2 x USB 3.0 ports, 1 x micro USB maintenance port
Video
1 x DisplayPort
Power & Dimensions
Power Supply
24 V (DC)
Thermal Solution
fan-less, via heat sink
Dimensions (W x D x H)
163 mm x 163 mm x 48 mm (without mounting plate)
163 mm x 210 mm x 48 mm (with mounting plate)