High-Performance Computing

At the core of Lickenbrock’s image processing software is high-performance computing (HPC) capabilities that allows for real-time evaluation of complex and computationally intensive solutions. The HPC solutions are made possible by massive parallel processing with GPUs and expertise in CUDA programming language. Performance is further improved with accelerated optimization techniques by numerical analysis of each unique problem.

Lickenbrock ports most of our developed prototype code onto GPUs for improved processing speed, and also contracts with other private companies to port their algorithms onto HPC platforms. Through the use of CUDA and code profiling, acceleration of some of the more compute intensive algorithms resulted in a speed up of 30x while maintaining accuracy. The software is compatible with both embedded and desktop GPUs.

Graphical User Interface Development

With all of the software and technology developed at Lickenbrock, corresponding graphical user interfaces (GUIs) and visualization tools are designed for a user friendly experience. The GUIs can be designed for both Windows and Linux computers.

3D viewer for CT inspection.

Computed Tomography (CT)

Lickenbrock is developing cone beam CT software that will automate the detection of defects in sophisticated aerospace components from CT scans. This includes a specialized 3D viewer, an iterative reconstruction algorithm, and an automated nondestructive evaluation system (ANDES). Our software package, Altux® CT, processes CT data to improve image fidelity. Altux® uses iterative image reconstruction to address beam hardening and scatter effects for both Fan Beam and Cone Beam CT. It produces fast high-resolution images and is compatible with most image acquisition software. Recent efforts involve developing compatibility with helical geometry.

Left: typical CT reconstruction. Right: CT reconstruction with scatter and beam hardening removal.

Image Restoration for Overhead Aerial Imaging

Overhead imagery is acquired via airborne platforms in a variety of applications. This imagery subject to multiple distortions and noise due to atmospheric disturbance, turbulence, and transmission defects such as compression artifacts and dropout. Lickenbrock has developed software that restores this imagery, which includes de-blocking and de-compression, image sharpening, super-resolution, contrast enhancement, and general denoising and deconvolution,  returning high-quality restored frames.

Left: corrupted overhead image. Right: restored image.

Left: image reconstructed from conventional synthetic aperture radar (SAR). Middle: a denoising method developed by Lickenbrock that removes noise but leaves some of the natural texture that may be desirable in the SAR image. Right: an alternative denoising method developed by Lickenbrock.

Multi-frame Blind Deconvolution for Ground Based Telescopes

Lickenbrock has developed a blind deconvolution approach that corrects blur in telescope images caused by atmospheric turbulence, optical aberrations, noise, and other factors. The solution interfaces with the imaging feed and corrects images in real time at typical video streaming rates. This solution improves images so that finer detail can be seen without requiring a telescope upgrade. This software also has the capability to correct for hot pixels and resolve at super-resolution.

Left: video acquired by ground-based telescope. Right: restored high-quality imagery.

(a) Deconvolution of an image containing a hot pixel. The red arrow in the middle indicates the location of the artifact resulting from the hot pixel, and the right image shows the result of the corrected deconvolution technique with an automated calibration. (b) Sub-pixel deconvolution of rocket body imagery.

Automated Terrain Classification

Lickenbrock is beginning to develop a fast classification algorithm to automatically label landscape terrain from overhead satellite and drone imagery. The algorithm works by computing local textures and features on the image and classifying each pixel based on these features using a trained classifier. An example of the performance of the prototype is shown below.