Hyperspectral imaging provides the power of viewing the same spatial scene under different electromagnetic spectra. Imaging sensors obtain scene images in spectral bands from visible-and-near-infrared (VNIR) to longwave-IR (LWIR) (0.2 - 12 microns). The absorption and reflectivity of different materials provide a unique "spectral signature," identifiable only at spectral bands beyond the visible range. Processing techniques generally identify and discriminate materials through these signatures.
Hyperspectral image processing has a wide variety of important applications including remote sensing to estimate agricultural crop yield, mineral detection, atmospheric cloud identification, military surveillance and tracking, manufacturing, and security. Real-time processing of hyperspectral data streams requires tremendous computational workloads and I/O throughput. Many of these systems will be deployed in a variety of platforms including vehicles, wearable computers, robots, and smart munitions. The key to successful deployment of these systems is the ability to extract and disseminate critical information from sensor data streams in a timely fashion. Because these systems are to operate in a mobile and perhaps covert manner, stringent resource limitations (size, weight, and power) pose additional system design challenges.
This research investigates focal plane parallel architectures as an efficient computational solution for real-time hyperspectral image processing. In particular, we process hyper-spectral data streams on the SIMD Pixel processor (SIMPil), which is a fine-grain parallel architecture developed at Georgia Tech for focal plane image processing. By employing a large array of stream processors with parallel interconnect between image sensors and processors, the architecture alleviates data bandwidth requirements, allowing computation to be performed while data arrives from the sensors in a stream-parallel computation model. Unlike traditional SIMD systems, SIMPil maintains high performance and modest generality in a low power and portable environment.
As part of this research, several key hyperspectral image processing applications, such as region autofocus, C-means classifier, K-means clustering, vector quantization image compression, and textural correlation using discrete Fourier transform have been studied. Critical information on computation workload, data throughputs, and memory storage are investigated for each application.
Publications: