Information Technology : Image processing


Energy-Efficient Multiplier Circuitry for GPUs

UW–Madison researchers have developed a new circuit system for multiplying floating-point numbers. The system combines a traditional floating-point multiplier with a power-of-two multiplier that works by shifting operations. Substantial power savings may be realized by selectively steering some operands to the power-of-two multiplier.

The different circuits have different advantages. The floating-point multiplier uses more power but is more versatile, while the power-of-two multiplier uses less power but is less versatile.

Statistical Imaging Reconstruction Is Faster, Cuts Noise

A UW–Madison researchers has developed an iterative reconstruction method that simultaneously achieves high convergence speed and high parallelizability. The method can work with various medical imaging systems, including CT, magnetic resonance imaging (MRI), X-ray angiography and positron emission tomography (PET).

In general, a nonlinear reconstruction problem is decomposed into separate linear sub-problems that can be solved more efficiently. The statistical image reconstruction process is decomposed into a statistically weighted algebraic reconstruction update sequence. After this step, the image is de-noised using a regularization function.

Simultaneous Image Reconstruction and Artifact Reduction

A UW–Madison researcher has developed a system for reconstructing images with different levels of artifacts. In this way, a ‘target image’ with the lowest level of problems will be produced simultaneously with an ‘artifact image’ that depicts primarily artifacts.

The method works by automatically and iteratively producing multiple images from one set of data, with the multiple images corresponding to different data consistency levels.

Once a subject is scanned, an image matrix is initialized having columns that correspond to different images. At least one image then is reconstructed by minimizing a matrix rank. The ranking is constrained according to a consistency condition that promotes the forward projection of each column to be consistent with a different subset of the acquired data.

Statistical Noise Map for Reducing X-Ray Exposure

UW–Madison researchers have developed a system and method for estimating a statistical noise map from a single X-ray exposure. This map accounts for noise acquired with X-ray imaging systems, including computed tomography (CT), tomosynthesis and C-arm systems.

The method reconstructs an image from acquired data using any standard filtered back projection (FBP) algorithm. This image is used as a baseline to estimate a noise standard distribution map. The raw projection data represents a typical measurement among many repeated measurements under the same experimental conditions. Therefore, this data can be used to generate several (e.g., 20 or more) noisy data sets.

These data sets are used to reconstruct noisy images that can be subtracted from the original image, resulting in a statistical noise map. This map accounts for a physical model of noise.

Efficient Video Retargeting Approach That Avoids Jitter

UW–Madison researchers have developed a video retargeting approach that provides both efficiency and temporal coherence. This method warps specific areas of each frame independently, yet avoids introducing jitter. Like previous approaches, this method warps frames so that background regions are distorted similarly to prior frames while avoiding distortion of the moving objects. In contrast to previous methods, this approach introduces a motion history map that propagates information about the moving objects between frames, allowing for graceful tradeoffs between temporal coherence in the background regions and shape preservation of the moving objects.

Universal Signal-to-Noise Ratio Enhancement Using PICCS Image Reconstruction

UW–Madison researchers have developed a universal method to improve SNR of a digital signal or image, including images produced using any medical imaging modality. The new method implements Dr. Chen’s previous discovery known as “PICCS” (see WARF reference number P08127US), which allows a high quality image to be reconstructed from undersampled image data. A final image with high SNR is constructed by imparting the high SNR characteristics of a “prior image” to the target image. This prior image is created from the original image, which allows an image to be improved without actually obtaining a prior image from the patient.

The first step in the new method is re-sampling, which converts a digital signal or image into a different domain that can be inversed easily for reconstruction purposes. These domains include radon, x-ray, Fourier or wavelet transform. Next, a filter is applied to the re-sampled signal or image to generate a very low noise prior image with low spatial resolution. Then, the PICCS algorithm is applied using the prior image to reconstruct the target signal or image. The resulting final image will have similar noise characteristics as the low-noise prior image, but the degraded spatial resolution will be restored in the iterative image reconstruction procedure.

Robust, Efficient and Streamable Video Stabilization

UW–Madison researchers have developed a robust and efficient approach to video stabilization that achieves high-quality camera motion for a wide range of videos. The key to this approach is that they enforce subspace constraints on feature trajectories while smoothing them. The method assembles tracked features in the video into a trajectory matrix, factors it into two low-rank matrices and performs filtering or curve fitting in a low-dimensional linear space. 

Night Vision System, Device and Method for Enhanced Signal-to-Noise Ratio

UW-Madison researchers have developed a method for displaying images using motion adaptive frame integration with real time digital processing and display. The method for filtering a series of image frames of a moving subject improves the signal-to-noise ratio of each image frame. The method combined with an optical apparatus configured to receive light from the scene comprises a lightweight, non-intensified imaging system for night vision.

A filtered image is formed by combining pixel values in a current image frame with weighted pixel values in previous and subsequent processed frames. Motion of the imaging system is compensated for so that the corresponding pixels in each image frame are registered to the same pixel locations before filtering. In effect, this registers the series of image frames with each other, providing an enhanced signal-to-noise ratio.

Algorithm for Estimating Parameters from Phase Data Without Unwrapping for Studying Earth’s Surface

UW-Madison researchers have developed an algorithm for interpreting an interferogram without the need for unwrapping. To do so, the invention interprets the interferogram by estimating parameters in a quantitative model directly from the wrapped phase data. Alternative unwrapping algorithms have been developed, but these can provide inadequate results in areas where the phase data are imperfect, leading to errors in the unwrapped phase values. Likewise, these algorithms rarely, if ever, provide uncertainty estimates, limiting attempts to weight the data in statistical analysis. Implementation of the invention would reduce the time and resources necessary for advanced interpretation of InSAR data products, and would provide a more accurate result that includes an assessment of the uncertainties of the parameter estimates.

Method for Assessing 3-D Crystal Structures from in Situ Digital Images

UW–Madison researchers now have developed a 3-D image analysis method to automatically extract information from developing crystal populations. The method is based on object recognition and can extract crystal size and shape distributions from low-quality in situ images in which the particles are overlapping, out of focus, randomly oriented or poorly illuminated. 

The method segments, or separates, objects from the background portion of the image. It uses a 3-D wireframe model to more accurately segment each particle.  Once the objects are removed from the image, shape, size, orientation and other relevant information can be determined for individual particles.  This information then can be used to control processes associated with the crystals, maximizing manufacturing efficiency.

Methods for Displaying a Large Image on a Small Screen Using Fish Eye Warping

UW-Madison researchers have developed a method for automatically adapting a large screen image, such as one taken by a digital camera, to a small screen, such as on a cellular phone or PDA. This method retains the context of the original image while emphasizing the information content of a region of interest from the image.

An image processing method is used to identify the important parts of the image. Then fisheye image warping methods are used to alter the image so the most important part is emphasized. Regions of the image surrounding the region of interest are warped to fit into available display space without regard to preserving their information content and/or aspect ratios, while the region of interest is modified to preserve its aspect ratio and information content.

Optimizing Probes to Improve Spectroscopic Measurement in Turbid Media

UW-Madison researchers have developed a method, apparatus and corresponding computer program to determine the optimal probe geometry for use in a particular tissue or other turbid medium. Light transport through the medium is modeled to simulate the diffuse reflectance properties that would be measured by a specific probe geometry. An inversion algorithm then converts the diffuse reflectance properties into optical properties. Those optical properties are compared to the known optical properties of the tissue to determine how well they match. When this process is repeated for additional probe geometries, the program indicates which geometry gives the most accurate measurement of the optical properties of the medium.

Systems and Methods for Recognizing Objects in an Image

UW-Madison researchers have developed a novel summation invariant recognition system for identifying objects with salient boundaries, such as fish or airplanes. This system represents the shape of an object as a set of discrete points and uses these points to create a summation invariant, which does not change under various viewing conditions.

To identify an object in an image, a contour of the object is extracted and normalized. Then one or more summation invariants are determined and compared to summation invariants for the target object(s). To distinguish between similar objects, the algorithm also can be performed locally to extract features of the contour and determine semi-local summation invariants. If the summation invariants match, the extracted object is recognized as the target object.

Systems and Methods for Automatically Determining Object Information

UW-Madison researchers have developed an algorithm designed to automatically extract information about crystal size from in situ digital images of suspended, high-aspect-ratio crystals. The algorithm combines linear features within the digital images into related groups and analyzes them to determine information about the size, shape and orientation of the crystals. This information can be used to control processes or devices associated with the crystals, thereby maximizing manufacturing efficiency.

Improved Method for More Efficient Mapping of Images

UW–Madison researchers have developed a computerized apparatus and an associated method and program for producing a flattening map from a digitized image. The digitized image may be a real object, quasi-discrete data or computer-generated discrete data. A first set of data comprising a plurality of discrete surface-elements is constructed to represent at least a portion of the surface of a digitized image; then, a flattening function is performed on the data set to produce the flattening map. The flattening map is conformally mapped onto a computer-generated surface that then can be displayed on a computer-assisted display apparatus that is in communication with a processor.

Method and Apparatus for Eliminating Ringing Artifacts in Decompressed Electronic Images

UW-Madison researchers have developed an efficient post-processing algorithm and software package that use mathematical morphology to significantly reduce ringing artifacts in highly compressed images. The method first identifies the edges in a decompressed image, and then defines zones around the edges where ringing artifacts are likely to occur. It then applies morphological filtering operators specifically to these zones to remove or reduce ringing artifacts in the final decompressed image.

Photorealistic Three-Dimensional Models of Real Scenes by Voxel Coloring

UW–Madison researchers have developed a new scene reconstruction algorithm that guarantees a consistent reconstruction is found, even under severe visibility changes. The technique generates a photorealistic three-dimensional projection of a real object or scene by modeling intrinsic color and texture information.

Specifically, the voxel coloring algorithm works by discretizing scene space into a set of voxels that are traversed and colored in a special order. It identifies a special set of invariant voxels, which together form a spatial and photometric reconstruction of the scene, fully consistent with the input images.

Improved Image Compression System Using Block Transforms and Tree-Type Coefficient Truncation

UW–Madison researchers have developed an apparatus and a method for compressing the size of a digitized signal. The method combines the benefits of block-wise processing with tree-type compression. The apparatus comprises a block selector, frequency analyzer, tree sequencer and coefficient truncator and may further include a quantizer and an entropy encoder. The researchers also include an apparatus for decompressing a compressed digitized signal.