Capitalizing on Today’s Computer Vision‒Based Quality Control Opportunities
Product quality monitoring solutions built on AI computer vision are poised to deliver tangible results to manufacturers and retailers. Using cameras, compute, and AI deployed on-site in stores and factories—or at the edge where visual data about products is collected—companies can monitor and assess product quality in near-real time. Regardless of the use case, whether inspecting products coming off an assembly line for defects or assessing strawberries in the produce section for freshness, many organizations struggle to get started with computer vision, and they anticipate further cost and scalability challenges as they move from proof of concept to production.
Ultimately, different types of organizations will take different routes to AI-enhanced product quality monitoring capabilities. Enterprise retailers and manufacturers are likely building AI capabilities in house, while small and midsize companies in those industries will more commonly look to independent software vendors (ISVs) and solution integrators (SIs) for support. These vendors and solution providers are being asked to make computer vision‒enhanced AI capabilities a reality while working within tight cost restraints and performance requirements.
Whether you’re an ISV, SI, retailer, or manufacturer, let’s review some essential considerations that can help you roll out AI-enhanced product quality monitoring with optimized efficiency and speed.
Training AI Models for Computer Vision
Core to any computer vision‒based product quality control strategy is the use of AI models at the edge to perform inferencing. Models are first trained, prior to deployment, to specialize in detecting product defects or recognizing product expiration. They are then deployed on edge devices as part of the AI-enabled software that analyzes the data from cameras to identify problems and report issues back to facility staff so they can take action to solve the issue.
Whereas enterprise organizations in retail and manufacturing may be more inclined to train these models from scratch or fine-tune existing models using their in-house staff, small and midsized companies will likely opt to rely on their technology partners to help enable the AI capabilities they need. This could be as simple as purchasing a solution from an ISV designed for their specific use case. Or they could rely on an SI to stitch together several solutions and handle any model customization that needs to happen.
As you seek to enable computer vision solutions, it’s important to remember that you don’t need to start from scratch. Existing models—many freely available online—can serve as a starting point for customization, helping to accelerate your efforts. For example, the Open Model Zoo for OpenVINO™ toolkit provides a range of deep learning models that can be retrained or fine-tuned for target applications.
Hardware
From a hardware perspective, you’ll need different levels of compute, depending on the training task you’re undertaking.
If you’re starting with an existing model to meet your or your customers’ needs, you may be able to fine-tune or retrain with cost-effective, general-purpose hardware. Doing so can help you avoid overinvesting in specialized resources you don’t truly need while reducing overall architectural complexity.
Training complex models from scratch, however, could require the advanced performance of specialized accelerators that are purpose built for AI workloads. These technologies are designed from the ground up to deliver the extreme performance needed for complex model training with massive parameter sets.
For example, Intel® Gaudi™ AI accelerators are purpose built to support demanding training and inferencing AI workloads. For retraining and fine-tuning, Intel® Xeon Scalable processors with integrated AI engines are particularly well suited to helping you meet performance requirements with a CPU-only architecture.
Software
A key challenge for many organizations is translating their teams’ domain- or industry-specific expertise into usable AI solutions. A variety of software platforms are available to help solve this challenge. These tools can help dramatically accelerate time to value for computer-vision solution-development initiatives.
For example, the Intel® Geti platform enables industry experts to build production-ready AI models quickly and collaboratively in a single interface, with minimal data science expertise. Users can easily add images or video data, create annotations, and train, retrain, export, and optimize AI models for deployment.
Deploying AI Models to Edge Hardware for Computer Vision
Many organizations assume that deploying computer vision-based quality control always requires heavy-edge infrastructure equipped with GPUs. This assumption is not entirely true. Today’s manufacturers and retailers can use common hardware resources to power AI computer vision at the edge.
Modern CPUs are well equipped to handle computer vision workloads at the edge. As in training, relying on CPUs for edge inferencing prevents technology overinvestment, streamlines deployments, and helps optimize power efficiency. CPUs can easily be deployed in ready-to-use, small form factor ruggedized systems to support the wide variety of operating environments encountered at the edge.
Processors such as Intel® Xeon processors and Intel® Core™ Ultra processors offer power-efficient AI performance that’s ideal for edge computer vision workloads.
Many manufacturers and retailers are already running critical edge systems featuring CPUs, including point-of-sale systems and software-defined IT/OT infrastructure. In many cases, these existing investments can support computer vision at the edge while minimizing the need for additional hardware purchases.
If needed, GPU hardware can be employed to meet more-demanding performance requirements. GPU technologies typically involve a larger power and size footprint but enable edge deployments to support advanced capabilities and innovative use cases.
For example, we offer both the Intel® Arc GPU and Intel® Data Center GPU Flex to help enable advanced computer vision capabilities at the edge.
That said, hardware is only one side of the equation. ISVs and SIs need ways to help their retail and industrial customers more easily handle heterogeneous deployment environments, optimize performance, and simplify any required model development. Here, ISVs and SIs can leverage software resources to help expedite and streamline their efforts.
For example, the OpenVINO™ toolkit can help you optimize, fine-tune, and run comprehensive AI inference using an included model optimizer and runtime and development tools. It’s an open source toolkit that helps you accelerate AI inference with lower latency and higher throughput while maintaining accuracy, reducing model footprint, and optimizing hardware use—all essential ingredients for scaling AI computer vision across multiple factories or stores. The toolkit can also be used to then convert and optimize models trained using popular frameworks like TensorFlow and PyTorch.
Built on Intel® oneAPI, OpenVINO™ helps SIs and ISVs more easily deploy AI inferencing at the edge to support computer vision solutions. By using the OpenVINO™ toolkit, you help ensure that your software offerings can run on the wide range of hardware that retailers and manufacturers may already have deployed. It also helps simplify technical requirements for greenfield deployments.
This kind of flexibility helps enable scalable, lightweight edge solutions that can deliver near-real-time insights without relying on cloud infrastructure. In doing so, these on-premises solutions avoid the cost, complexity, and security concerns associated with transporting data to the cloud for processing.
Start Capturing the Potential Today
For ISVs, SIs, and the retailers and manufacturers they serve, now’s the time to adopt a scalable and efficient approach to AI-enabled product quality control. Intel and our ecosystem of partners are committed to delivering the open, interoperable technologies you need.
As you continue exploring the solution development possibilities available to you, keep in mind that you can try many of the hardware and software technologies discussed in this article for yourself via the Intel® Tiber™ Developer Cloud