content provided by pleora

As system designers and integrators navigate through Industry 4.0, Internet of Things (IoT), and artificial intelligence (AI), it is clear that they must balance the best solution for the problem and plan for integration with existing infrastructure and processes. End-users, including brand owners, quality managers, and system integrators evaluating AI consistently highlight a few key factors that potentially limit their ability to deploy AI. 

1. Lack of Technology Infrastructure to Support AI

This is a key concern, in particular for manufacturers with a significant investment in inspection infrastructure. A major automotive manufacturer, for example, has over 2000 cameras completing 2 million inspection tasks per day on just one production line. Many also want to avoid “vendor lock-in” as they consider infrastructure and system investments.   AI needs to complement, not replace, existing infrastructure. 

2. Lack of Available Collected Data

Costs related to developing or acquiring specialized skills or consulting expenses for developed AI algorithms are typically highlighted by end-users considering deployment scenarios. Training AI for deployment requires better access to data and easier to develop algorithms. 

3. Proven Frontline Processes

In a visual inspection application there are both human and software processes delivering proven results. The goal of AI is often to enhance decision-making, not necessarily replace it. End-users are also concerned with retraining costs for AI platforms compared to existing software applications. AI needs to adapt to how employees currently work, versus adapting employees to how AI works.


What is AI, Machine Learning, and Deep Learning? 

AI is undoubtedly one of the most hyped technologies of recent times, and with market hype comes a lot of confusion.

AI is broadly defined as a machine solving a problem or completing a task in a way that we consider “intelligent”, like solving a math problem. AI is a very broad category, and includes applications such as natural language processing, speech recognition, and eCommerce. 

Machine learning is a branch of AI where machines learn how to solve a specific problem without human intervention, so long as they are supplied with data. Deep learning is a branch of machine learning that refers to neural networks, which are composed of artificial neurons inspired by how the human brain works. Neural networks are characterized by the ability of the network to “learn” the features relevant to the problem being solved as it sees more and more data. 

A neural network is able to improve its performance and “teach itself” as it collects data, like we do as humans (a child learning how to ride a bike becomes better and better each time they try). As the neural network has to make a decision, it will analyze all of the data available and make a prediction. The more decisions the network makes, the more it learns to correct itself via validation and confirmation. Over time the machine improves its abilities and prediction skills.

AI has an end-goal of being autonomous. Machines may not immediately be fully autonomous, but through training the goal is to develop AI capabilities that can function and process decision-making skills independent of human intervention. In an inspection application, for example, the goal is to be able to train the system to the point where it can begin learning errors and independently identifying a defective product. 

There is a significant amount of research now targeted towards machine vision capabilities, such as object classification, detection, and segmentation. By adding AI skills alongside traditional vision processes, manufacturers can significantly reduce quality inspection errors that cost money in terms of human capital and scrap product. For a large manufacturer, false positives are a significant issue that result in product line downtime and costly secondary human inspection. Classification with a simple two classes — pass/fail — is the quickest and least complex application to train when starting out with AI.


Hybrid AI Offers a Solution 

To understand the benefits of hybrid AI, let’s revisit the common end-user and integrator concerns while considering a traditional vision inspection application. New hybrid AI approaches integrate algorithm development and edge processing to provide users and integrators with a flexible end-to-end solution.  

AI Algorithm Development and Training

In a classic computer vision application, a developer manually tunes an algorithm for a task to be completed. This can require significant customization if products A and B have different thresholds on what is considered an error. Inaccuracies may generate excessive false positives that stop production and force costly manual secondary inspection, or missed errors that result in defective or poor quality products going to market.

Similarly, AI algorithm training has traditionally required multiple time-consuming steps and dedicated coding to input images, label defects, fine-tune detection, and optimize models. More recently, companies are developing no-code software platforms that provide an intuitive drag-and-drop approach to develop “plug-and-play” machine learning quality inspection applications. Users can design and deploy advanced computer vision, AI, machine learning and deep learning capabilities in minutes instead of days, without requiring specialized skills or external consulting. 

AI plug-ins can be developed in any standard web browser on any device with no-code “block-based” tools for vision and AI programming. Comprehensive platforms allow the design of traditional computer vision plug-ins for standard features (for example detection, thresholding, measurement, pattern matching, and barcode reading) and deep learning classification and object detection skills. For more advanced users, platforms provide full flexibility for developers to code and test plug-ins using Python, customize computer vision and AI plug-ins, and create custom plug-ins to run any type of AI model. In addition, “off-the-shelf” plug-in AI skills for quality inspection and hyperspectral imaging lets users easily deploy advanced capabilities for common requirements.  


Edge Processing 

A hybrid approach takes advantage of advances in edge processing and embedded technologies to seamlessly add AI capabilities alongside existing infrastructure. Plug-ins are transferred to the embedded platform, which acts as an intermediate device between the camera and host PC. The embedded device “mimics” the camera for existing applications and automatically acquires the images and applies the required AI skills. Processed data is then sent over GigE Vision to the inspection application, which receives it as if it were still connected directly to the camera. The embedded device can also transfer images back to the software platform for continuous offline training. 

In a potential scenario, a brand owner could begin with a hybrid approach to deploy offline inspection to automate a visual inspection process. For example, AI can enhance human processes by flagging images or products for operator analysis. The device could also be used as a secondary inspection tool by processing imaging data with loaded plug-in skills in parallel to traditional processing tools. If a defect is detected, processed video from the embedded device can confirm or reject results as a secondary inspection.

While continuously training the system with collected data, and building confidence in results, the end-user can gradually transition to an AI-based inspection model. Over time, the end-user can begin to use inspection data as part of a more comprehensive analytics-based Industry 4.0 initiative, focused on driving efficiencies. This can include using data to monitor machine performance for proactive maintenance, and leveraging cloud capabilities to share data across global facilities to improve inspection processes.  


Where is Hybrid AI Deployed Now?

Hybrid AI is now being used to improve efficiencies, quality, and profits across a wide range of markets, including consumer goods, automotive, food & beverage, and print & packaging.

Consumer Goods: AI is being used to enhance traditional machine vision quality inspection reduce false-positives and production errors that typically require costly secondary human inspection. This includes screening for defective goods, as well as “brand management” inspection to ensure goods are visually appealing to consumers. 

Automotive: Hybrid AI is being used to inspect components within major systems of automobiles, including engine, drive, braking and electrical systems. This can include error proofing, dimensional verification, and surface quality processes that often rely on a combination of traditional vision and human inspection. 

Food and Beverage: Like consumer goods, hybrid AI is enhancing processes to ensure quality and brand management. Food inspection, for example, is adopting hyperspectral imaging to detect foreign materials and ensure products meet quality standards while reducing costly visual inspection. Many software processing solutions require custom workarounds to support hyperspectral through GigE Vision because they can’t interpret hyperspectral multiband information. In comparison, hybrid AI bridges the gap between applications and existing machine vision software by automatically handling image acquisition from the hyperspectral imaging source and sending out the processed data over GigE Vision to inspection and analysis platforms.