While innovation in computer vision largely comes from computer science fields, this study has also been greatly facilitated by improvements in hardware. EEs are continually coming up with new methods of imaging and sensing, providing computer vision algorithms with more useful raw data.
One of the developing technologies that have found great use in scientific and industrial computer vision applications is short-wave infrared (SWIR) imaging. Recently, research institution Imec touted that its newly-developed photodetector has achieved the smallest pixel pitch to date with SWIR technologies.
SWIR images of three separate pixel pitches. Image used courtesy of Imec
To better understand the significance of this announcement, it may be helpful to first assess how SWIR imaging functions—especially as this technology becomes increasingly useful in future machine learning applications.
How SWIR Works
SWIR imaging utilizes the 1-2.7μm wavelength region of the infrared spectrum.
SWIR-based sensors operate on a similar principle to CMOS-based sensors, converting photons into electrons. Unlike CMOS sensors, which are based on silicon, SWIR technology must use either indium gallium arsenide (InGaAs) or mercury cadmium telluride (MCT or HgCdTe), which are both sensitive to infrared light.
These sensors’ sensitivity to different wavelengths is a function of their chemical structure.
InGaAs sensor architecture. Image used courtesy of Photonics Media
Both materials require strong cooling to achieve a proper signal-to-noise ratio (SNR), but MCT often needs cryogenic cooling. As a result of this, InGaAs is more commonly used because it’s more practical and affordable. These types of sensors work with a hybrid architecture that combines a silicon-based CMOS readout circuit with an InGaAs photosensitive array.
SWIR opens doors to enhance current machine vision systems by imaging beyond the visible spectrum.
Due to its reflective nature, SWIR light contains shadows and contrast in its imagery. Producing colorless images, SWIR-based imaging makes objects easily recognizable, providing huge enhancements to computer vision applications. SWIR also allows cameras to see through surfaces that are not transparent to the human eye.
Filled bottles imaged with visible light (left) and SWIR (right). Image used courtesy of Photonics Media
One classic example in which SWIR is useful is in moisture detection. While water is transparent to visible light, it absorbs SWIR, which makes it appear identifiably black in a resulting image. Other applications include non-invasive machine inspection and spectroscopic analysis. Even tasks like plastics sorting for recycling are made possible thanks to the advantages of SWIR.
Imec’s SWIR Sensor Offers Record Small Pixel Pitch
Further pushing the development of SWIR technologies, Imec made headlines this week with news of a prototype monolithic SWIR photosensor that achieves record resolution.
Conventional SWIR image sensors are generally produced with hybrid technology that flip-chip bonds an InGaAs-based photodetector to a silicon readout circuit. While achieving high sensitivity, this setup can be expensive to manufacture and is limited in pixel size and density, impeding the adoption of this technology.
Imec’s photodetector integrated on a CMOS readout circuit. Image used courtesy of Imec
Imec’s new prototype circuit is instead based on a Si-CMOS thin-film photodetector that is monolithically integrated. Imec says this circuit is “fab compatible” for high-throughput wafer-level manufacturing.
Based on 130nm CMOS technology, the prototype works using a thin absorber layer such as 5.5nm PbS quantum dots, which corresponds to peak absorption at 1400nm wavelength. This peak absorption, however, can be tuned to wavelengths even greater than 2000nm
With this new approach, the photodetector is said to achieve a pitch of 1.82μm—a feat Imec claims is the record smallest pixel pitch achieved with SWIR technologies to date.
Accurate SWIR Yields Accurate AI
EEs who work with SWIR imaging technology have an important role to play in computer vision innovation. This technology can capture unique images beyond the scope of CCD or CMOS image sensors.
As SWIR imaging devices become even more precise (just as Imec is striving to achieve), machine learning algorithms will serve up more detailed and accurate results across a number of fields: assessing water content in plants, detecting moisture in woodworking, and even identifying pencil or charcoal under historical paintings.