SBIR-STTR Award

Rice in concert with Inview Technologies will design, simulate, and evaluate compressive algorithms for object detection and classification in both static and dynamic imaging. The eventual goal is the application of the most suitable methods in novel short-wave and mid-wave infrared optical architectures for high-speed discovery and tracking by exploiting sparse signatures. We will compare our newly developed methods against state-of-the-art approaches such as the principal component and linear discriminate methods currently in use with traditional focal plane array imagers. We will utilize our experience compressive video imaging to design new mathematical projections to identify and isolate foreground targets from cluttered background. We will begin from our Hadamard approximations of both for video reconstruction but will further refine the compressive domain separation by exploiting machine vision algorithms such as our manifold secant approach as a means to improve signal-to-noise performance of the receiver operating characterstics. The second approach will build up from our successful use of Inview’s Partial Complete methodology of building up anomaly detection patterns from select Hadamard kernels. While very generic, these exploit localized commonalities among the targets and will be further expanded to an optimal multiscale collection of patterns that accurately reflect the target’s individualistic features.

Benefits:
We aim to show that compressive domain algorithms and processing techniques can provide actionable decision making capabilities to platforms such as weapons seekers that have limited processing power.

Keywords:
Compressive sensing, anomaly detection, compressed domain processing, machine learning, target recognition, target classification

InView Technology Corporation and its partner Rice University propose the simulation, analysis, development and demonstration of a novel infrared imaging system that provides enhanced compressive sensing-based target detection, classification and tracking capabilities where all processing and reasoning techniques are performed directly in the measurement domain, without the expense of sparsity-based signal reconstruction. Performing signal processing directly in the sparse measurement domain has several advantages including the elimination of signal reconstruction time and processing costs and a reduction in the volume of raw data that limits computational and communication throughput. In computer vision applications not only is there no need to reconstruct signals or images, but autonomous machine learning can proceed with higher efficiency in lower dimensional spaces. In Phase I, feasibility and robustness of such an approach was demonstrated using simulation and further evaluated the performance of various the computationally-based optical approaches to performing compressive sensing-based target detection directly in the compressed domain on a short-wave image target library. The objective of Phase II is to expand, optimize and benchmark these mathematical techniques and implement them on an operational infrared compressive sensing hardware platform both in the laboratory and the field.

Benefits:
The expected outcome of this Phase II project is a prototype-level demonstration of an infrared compressive imaging system that can perform machine vision tasks more efficiently than traditional image processing algorithms acting on focal plane array measurements. The compressed domain approach to target detection, classification and tracking has applications in electro-optical imaging sensors for weapon seekers, persistent surveillance systems, standoff detection of chemical and biological threats, autonomous vehicle navigation, and air-to-ground weapon applications where background clutter complicates the recognition of targets. Commercially, InView will target a $1B market for advanced infrared security, surveillance and navigation cameras, for high value installations: such as refineries, factories, oil platforms, Commercial ships and yachts and in Intelligence and Law Enforcement. Machine vision also plays a major role in autonomous vehicle navigation where high-resolution imaging must combine with rapid analysis and decision making. The prohibitive cost of focal plane arrays in infrared portion of the spectrum, has meant that machine vision is almost exclusively carried out in the visible spectrum using silicon-based imagers. The success of this project will allow such tasks to be implemented across the short- and mid-wave infrared portions of the spectrum in a much more affordable manner.

Keywords:
compressive sensing, automated target recognition, classification, machine vision, compressed domain image processing, neural networks, shortwave infrared, deep learning