Image Acquisition – The Linchpin of AI-Powered Machine Vision Systems

By August 22, 2022December 25th, 2022Image Acquisition
Image Acquisition

Advanced industrial automation technologies such as machine vision strive to replace human visual sensing and judgment capabilities with a camera, industrial computer, and software, with the objective of providing automated image-based inspection or analysis. Over the last two decades, machine vision systems have evolved rapidly to become an integral part of manufacturing industries solving a range of myriad tasks and helping in achieving superlative quality, while increasing throughput and efficiency. The frenzied rise in demand for machine vision in recent years has been made possible by the easy availability of powerful vision systems and the emergence of AI technologies in this domain. As we shall see subsequently, the machine vision subsystems have to perform synergistically to deliver the desired results, however, this article aims to highlight the significance of image acquisition and how it holds the key to the overall success in the entire gamut of machine vision.

Machine Vision System – An Overview

To enable a machine vision system to ‘see’, it progresses through three basic steps – image acquisition, image processing, and post-process automation. Image acquisition (IA) is the first step where the image of an object is acquired via a setup that includes a trigger element, illumination, camera, optics, and material handling. Typically, the first step in the IA stage is to find the object or feature of interest within the field of view of the camera. Lighting illuminates the part to be inspected and delivers the desired contrast, making the features clearly visible to the camera. The lens captures the image and presents it to the sensor in the form of light (optical image). A digital output of this optical image, which includes details such as color, brightness, intensity, and light scatter, is created and read off the sensor for further processing. (FIG: 1)

FIG:1– Setup for Image Acquisition

The next step, image processing is a method of performing specific operations on an image such as removing noise, changing the scale of the image, enhancing contrast and extracting specific features like lines, edges, points, or textures. In image processing, the input is the digital image acquired from the first step, and the output can be a modified image, or information defining the features associated with that image.

Post image processing, the processed data is used to take appropriate actions or make decisions. This can involve a rejection of the part, triggering a response to stop the line, manual intervention of some kind or sending a feedback or signaling to upstream or downstream processes for taking requisite actions.

Artificial Intelligence and Deep Learning in Machine Vision

Artificial Intelligence (AI), especially in the form of Deep Learning (DL), finds an important application in industrial machine vision. Developers are increasingly applying neural networks and their variants to improve object detection, classification, or addressing other problems through AI-powered vision systems that are difficult or nearly impossible to handle with traditional rule-based machine vision algorithms. DL techniques are capable of evaluating large datasets. AI-powered MV systems not only ensure the detection of defects efficiently and at a faster rate, but are also easier to tweak for new variants of a product when compared to conventional rule-based MV systems. Identification of the tiniest of the scratches, minute paint defects, poor food quality, and many such tasks, with capabilities outperforming the best of quality inspectors, have driven businesses to adopt AI-powered MV systems.

Combining AI with MV helps in automating the workflows and interventions that were once manual, by not only classifying what is ‘seen’, but also making the MV system smart enough to make decisions, thus imitating the capabilities of a human brain.

AI-Powered Machine Vision Systems and Image Acquisition

The Deep Learning algorithms that make Machine Vision Systems smart are extremely data-hungry. They require a significant amount of image datasets to fine-tune the large number of parameters that are estimated during inferencing.

A typical AI-based defect detection process acquires several images of the product, both good and defective, and then identifies and annotates them accordingly.  The corresponding DL model of the object is trained with this labeled data and subsequently tested with a set of images. If it fails to detect defective parts, changes are made in the settings or the images, and the process is repeated.

Once successful, the generated model is deployed for run-time execution and inferencing. If the DL model is still unable to perform and precisely detect the defect, one of the likely reasons could be the inferior quality of the images it was trained with.

To enable the model to derive the right inferences and minimize the false negatives and false positives during a quality inspection on the production line, the IA process must deliver reliable images.

The success or failure in capturing an accurate, good-quality image of an object is undeniably determined by the efficiency of the image acquisition process.

An ideal combination of all components of IA has to be arrived at by performing a detailed analysis of the requirements in terms of the product, inspection criteria, and factory dynamics. The different components (FIG: 2) and the factors and challenges associated with each of these will be discussed in detail in an upcoming series of articles.

Sharpening the ‘vision’ is the mission

DL models train on large amounts of relevant and quality data to create a real-world, high fidelity application. Building such datasets is expensive both in terms of effort and cost. Moreover, no amount of image processing can improve the quality of the image if it was sub-standard at the image acquisition stage itself. Only a well designed image acquisition system with the right combination of matched components and software can help in building such datasets. Qualitas Technologies, with its know-how developed over a decade, has a complete solution in the Qualitas EagleEye®Suite.

How do you like our blog? Let us know in the comment box. Also, feel free to ask anything related to the same topic.

Leave a Reply

Schedule A Demo
close slider

hbspt.forms.create({ region: "na1", portalId: "726123", formId: "bd9f048a-7e04-4512-9179-c50855c961ea" });