A project from apparently about 3 days ago on Edge Impulse.
Full details in link.
An industrial inspection application that uses the Brainchip Akida Neuromorphic processor for fast and efficient quality control inferencing.
docs.edgeimpulse.com
Industrial Inspection Line - Brainchip Akida Neuromorphic Processor
An industrial inspection application that uses the Brainchip Akida Neuromorphic processor for fast and efficient quality control inferencing.
Created By: Peter Ing
Public Project Links:
Object Detection -
https://studio.edgeimpulse.com/studio/349843 Classification -
https://studio.edgeimpulse.com/studio/349858
GitHub Repo:
https://github.com/peteing/brainchip_edgeimpulse_inspectionsystem.git
Introduction
In the ever-evolving landscape of modern manufacturing, the efficiency and accuracy of production lines are paramount. The meticulous inspection of products at various stages ensures not only the adherence to quality standards but also the optimization of resources. In this dynamic scenario, the integration of cutting-edge technologies such as computer vision and artificial intelligence has emerged as a game-changer.
Initially, machine vision systems relied on basic image processing techniques and rule-based algorithms. These early systems were capable of performing relatively simple tasks, such as inspecting products for basic defects or checking for the presence of specific features. These systems required cameras with high-cost Industrial PC's to perform CPU based processing that was expensive and power hungry while offering limited performance.
Today the trend has shifted towards using Deep Learning specifically Convolutional Neural Networks on Graphics Processing Units and specialized CNN hardware accelerators. The solutions on the market are still relatively costly and power hungry. Camera and IPC's are available with integrated acceleration built-in for industrial use cases, but are very expensive.
Neuromorphic processing, inspired by the human brain, diverges from traditional computing with its parallel, adaptive features like Spiking Neural Networks, parallel processing, event-driven computation, and synaptic plasticity. This disruptive technology holds promise for energy-efficient, brain-like information processing, particularly in tasks like pattern recognition and sensory processing. This makes Neuromorphic computing ideal for use in Industrial Inspection systems where it can provide real-time insights into inspections. The benefits include reduced costs and improved performance and being able to adapt the system at the edge to new use cases.
Brainchip Akida represents the state of the art in production-ready Neuromorphic computing ideally suited to edge use-cases. We will be demonstrating the power of the Brainchip Akida in an industrial setting in this guide as part of a standalone inspection system that can be setup along a production line.
The Akida processor is available on a PCI-E card form-factor for integration into your own hardware, or ships as either an Intel or Arm-based developer kit. For the purpose of this project our focus is on the Arm-based developer kit, which consists of a Raspberry Pi Compute Module 4 mounted on a Raspberry PI Compute Module 4 IO board, which is what we are using for this application.
Many users coming from an Industrial environment have limited experience when it comes to AI and Deep Learning and this can seem daunting. There are very expensive platforms and solutions that help simplify the process, but none can match the ease of use and rapid performance of using Edge Impulse for the AI component of your project.
Industrial Inspection Use Case
A typical scenario in an industrial manufacturing plant is defect detection. This can be applied to a range of different product types but essentially the requirement is always to determine which products to reject, out of a set of products that are often in motion using some kind of materials handling equipment such as a conveyor.
To achieve this, classic machine vision techniques using old camera systems running CPU algorithms often included detecting a Region of Interest (ROI) and then focusing on that area, and using tools such as edge and blob detection to find anomalies.
Deep learning solves this approach by making use of learning algorithms to simply teach the system what is correct and what isn't. This results in a 2 stage pipeline that first does Object Detection, then cascades the results to a classifier.
The Object Detector functions as the Region of Interest segmenter, while the classifier then determines if a product is defective or damaged, or passes the quality check. We will proceed to implement such a pipeline together with a custom GUI based app.
