Embedded Vision
As deep learning capabilities integrate into a range of technology areas, engineers and designers must determine how-best to harness the potential of AI/DL for their systems, starting in the design and fabrication stages. Camera designers, as well as vision system integrators, must develop imagers and imaging systems capable of achieving optimal performance without compromising or restricting the AI/DL element. At the same time, cameras must support the AI element: Deep learning applications cannot achieve peak levels of functionality in poor lighting settings, or as part of a sub-optimally designed system, and require high-quality images for training and labeling via a data set. The considerations that camera designers and end-users give to system design in the budding AI age is explored. Considered questions include how designers are adapting camera designs to incorporate or pair with an AI/DL element; the relationship between AI/DL capabilities in image processing and the design of a camera; and the trade-off between the quality of camera optics and the sophistication of the AI element.
Key Technologies: 3D imaging, cameras, machine vision systems, camera optics, AI/DL, camera sensorss
Hyperspectral and Multispectral Imaging
"Imaging beyond the visible wavelength has become an increasingly popular tool in industries such as food and beverage processing, agriculture, and electronics and semiconductor manufacturing. Hyperspectral imaging, an advanced technique that captures hundreds of spectral bands across the ultraviolet, visible, near-infrared, and shortwave infrared spectrums, can identify and differentiate virtually any material. But the complexity of hyperspectral image data makes it impractical in production environments for several reasons, including specialized operation, high demands on processing power, and prohibitively high component costs.
In this article, Steve Kinney, director of engineering for Smart Vision Lights, explains how multispectral imaging — with accessible, familiar machine vision tools such as multispectral lighting, narrowband filters, and conventional cameras — can translate the benefits of hyperspectral data to a production environment. The result is a narrower, more applicable data set for conventional image analysis, yet delivering much richer image data and enhanced contrast."
Key Technologies: hyperspectral imaging, multispectral imaging, IR, machine vision
Next-Gen UAV Hyperspectral Processing
"Over the past few decades, significant advancements in processing power relative to both weight and power consumption have dramatically increased the computational capabilities available on lightweight platforms such as UAVs. These improvements have enabled far more complex data processing tasks to be performed in real time, even in resource-constrained environments. HySpex, NORCE and ReSe are leveraging these advancements to develop an innovative, real-time solution, both hardware and software, for converting raw hyperspectral data into fully visualized 3D hypermesh, as well as application-specific mesh tailored to diverse needs.
One of the key challenges is properly co-registrating VNIR (Visible and Near-Infrared) and SWIR (Short-Wave Infrared) imaging spectrometers, which have distinct sampling point spread functions (SPSF). Aligning the SPSF of these two systems is essential for achieving high spectral fidelity at the pixel level. Our solution involves merging data from the VNIR and SWIR spectrometers to cover the full spectrum from 400nm to 2500nm, producing a unified hyperspectral dataset that maintains spectral fidelity across the entire spectral range.
Beyond the spectral merging challenge, another significant hurdle is the real-time generation of a Digital Surface Model (DSM) during UAV flights. This is crucial for precise 3D scene reconstruction, and advanced processing techniques have been developed to generate DSMs in real time, even under the demanding constraints of UAV operations. The DSM is represented as a mesh, this is needed for the ray tracing. Once the DSM is created, the ray-tracing engine leverages it to generate a detailed mesh of the hyperspectral data, this we call a hypermesh.
In addition to producing 3D radiance data, real-time atmospheric correction using applied algorithms ensure accurate reflectance data. This corrected reflectance data is then used to generate application-specific meshs also in real time. The entire process works on unrectified data, eliminating the need for rasterization, which often introduces huge data sets that are limited to 2D (or 2.5D) representation. By avoiding rasterization, high-quality 3D visualization is achieved with less than 5% data size overhead on the unrectified data."
Key Technologies: hyperspectral imaging, UAVs, VNIR, SWIR, digital surface models
3D Vision
For industrial applications, 3D vision advances promise greater capabilities and better performance. On the hardware side, CMOS sensor improvements like higher dynamic range faster frame rates ensure reliable image quality in diverse situations while global shutter technology eliminates artifacts. The result is precise depth measurement in high-speed applications. On the software side, AI is a game changer, enabling self-learning, adaptive systems that improve robotic guidance, inspections, and automation. The next step is deep AI integration, which will benefit semiconductor manufacturing and other demanding applications that require extreme accuracy. Getting to that point, though, will depend upon technological innovations, the deployment of standards, and other enhancements required for success in an industrial setting.
Key Technologies: 3D vision; CMOS sensors; robotics, automation
Download Media Planner