We present a method for extracting depth information from a rectified image pair. Our approach focuses on the first stage of many stereo algorithms: the matching cost computation. We approach the problem by learning a...
详细信息
We present a method for extracting depth information from a rectified image pair. Our approach focuses on the first stage of many stereo algorithms: the matching cost computation. We approach the problem by learning a similarity measure on small image patches using a convolutional neural network. Training is carried out in a supervised manner by constructing a binary classification data set with examples of similar and dissimilar pairs of patches. We examine two network architectures for this task: one tuned for speed, the other for accuracy. The output of the convolutional neural network is used to initialize the stereo matching cost. A series of post-processing steps follow: cross-based cost aggregation, semiglobal matching, a left-right consistency check, subpixel enhancement, a median filter, and a bilateral filter. We evaluate our method on the KITTI 2012, KITTI 2015, and Middlebury stereo data sets and show that it outperforms other approaches on all three data sets.
Purpose: To quantitatively evaluate and compare the accuracy of two advanced methods that can estimate the level of noise per voxel in patient images. These noise estimation methods show promises in: 1) assuring the p...
详细信息
Purpose: To quantitatively evaluate and compare the accuracy of two advanced methods that can estimate the level of noise per voxel in patient images. These noise estimation methods show promises in: 1) assuring the performance of imaging systems and algorithms, 2) guiding imageprocessing tasks for clinical and research applications, i.e. by optimization of the parameters, and 3) quantifying patient image quality and assisting image quality improvements. Methods: We conducted an experiment of 34 repeated MRI scans (TrueFISP sequence) of a swine head in order to obtain a ground truth noise dataset. Two published noise estimation methods were implemented in this study: 1) Minimal Difference in Neighborhoods (MDiN) and 2) high-pass MDiN. Noise estimation accuracies of two methods were quantitatively measured using the ground truth data and patient MRI images with added Rician noise. Results: The experimental results with both swine head images and patient images showed that the MDiN method is more accurate. The high-pass MDiN method is slightly less but still sufficiently accurate. The MDiN method could be obtained within a 90% accuracy when tested on the ground-truth dataset. Conclusion: We verified the performance of two efficient methods to automatically estimate per voxel noise levels in patient images. Our results suggest that these methods could be confidently used to assist and guide clinical and research applications that require such noise information. Senior Author Dr. Deshan Yang received research funding form ViewRay and Varian.
Purpose: The daily treatment MRIs acquired on MR-IGRT systems, like diagnostic MRIs, suffer from intensity inhomogeneity issue, associated with B1 and B0 inhomogeneities. An improved homomorphic unsharp mask (HUM) fil...
详细信息
Purpose: The daily treatment MRIs acquired on MR-IGRT systems, like diagnostic MRIs, suffer from intensity inhomogeneity issue, associated with B1 and B0 inhomogeneities. An improved homomorphic unsharp mask (HUM) filtering method, automatic and robust body segmentation, and imaging field-of-view (FOV) detection methods were developed to compute the multiplicative slow-varying correction field and correct the intensity inhomogeneity. The goal is to improve and normalize the voxel intensity so that the images could be processed more accurately by quantitative methods (e.g., segmentation and registration) that require consistent image voxel intensity values. Methods: HUM methods have been widely used for years. A body mask is required, otherwise the body surface in the corrected image would be incorrectly bright due to the sudden intensity transition at the body surface. In this study, we developed an improved HUM-based correction method that includes three main components: 1) Robust body segmentation on the normalized image gradient map, 2) Robust FOV detection (needed for body segmentation) using region growing and morphologic filters, and 3) An effective implementation of HUM using repeated Gaussian convolution. Results: The proposed method was successfully tested on patient images of common anatomical sites (H/N, lung, abdomen and pelvis). Initial qualitative comparisons showed that this improved HUM method outperformed three recently published algorithms (FCM, LEMS, MICO) in both computation speed (by 50+ times) and robustness (in intermediate to severe inhomogeneity situations). Currently implemented in MATLAB, it takes 20 to 25 seconds to process a 3D MRI volume. Conclusion: Compared to more sophisticated MRI inhomogeneity correction algorithms, the improved HUM method is simple and effective. The inhomogeneity correction, body mask, and FOV detection methods developed in this study would be useful as preprocessing tools for many MRI-related research and clinic
Purpose: Iterative reconstruction algorithms in computed tomography (CT) require a fast method for computing the intersections between the photons’ trajectories and the object, also called ray-tracing or system matri...
详细信息
Purpose: Iterative reconstruction algorithms in computed tomography (CT) require a fast method for computing the intersections between the photons’ trajectories and the object, also called ray-tracing or system matrix computation. This work evaluates different ways to store the system matrix, aiming to reconstruct dense image grids in reasonable time. Methods: We propose an optimized implementation of the Siddon's algorithm using graphics processing units (GPUs) with a novel data storage scheme. The algorithm computes a part of the system matrix on demand, typically, for one projection angle. The proposed method was enhanced with accelerating options: storage of larger subsets of the system matrix, systematic reuse of data via geometric symmetries, an arithmetic-rich parallel code and code configuration via machine learning. It was tested on geometries mimicking a cone beam CT acquisition of a human head. To realistically assess the execution time, the ray-tracing routines were integrated into a regularized Poisson-based reconstruction algorithm. The proposed scheme was also compared to a different approach, where the system matrix is fully pre-computed and loaded at reconstruction time. Results: Fast ray-tracing of realistic acquisition geometries, which often lack spatial symmetry properties, was enabled via the proposed method. Ray-tracing interleaved with projection and backprojection operations required significant additional time. In most cases, ray-tracing was shown to use about 66 % of the total reconstruction time. In absolute terms, tracing times varied from 3.6 s to 7.5 min, depending on the problem size. The presence of geometrical symmetries allowed for non-negligible ray-tracing and reconstruction time reduction. Arithmetic-rich parallel code and machine learning permitted a modest reconstruction time reduction, in the order of 1 %. Conclusion: Partial system matrix storage permitted the reconstruction of higher 3D image grid sizes and larger projectio
One of the key problems in the field of Computer Vision is recovering the geometry from multiple views of the same scene. A feature-based approach to solve the challenge of finding matching points in different views i...
详细信息
ISBN:
(纸本)9781479937707
One of the key problems in the field of Computer Vision is recovering the geometry from multiple views of the same scene. A feature-based approach to solve the challenge of finding matching points in different views is the scale-invariant feature transform (SIFT). SIFT requires complex accelerated feature extraction combined with low energy requirements to meet the strict constraints of advanced driver assistance systems (ADAS) with regard to power consumption, processing speed and flexibility for future algorithms. This paper presents an application-specific instruction-set extension for a Tensilica Xtensa LX4 ASIP to accelerate a SIFT feature extraction and its evaluation. When compared to the same arithmetic functions processed on an ASIP without any extensions, basic elements of digital imageprocessing and specialized SIFT processing tasks that are accelerated reach a significant speed-up factor for arithmetic functions of x1300. At the same time the accuracy of the SIFT features is preserved. The SIFT feature extraction on an extended processor was accelerated by a factor of x167 compared to the base processor. In addition, the proposed processor extensions maintain the full flexibility of an ASIP for a fast integration of future feature extractors for advanced driver assistance systems.
Context. The planetary system discovered around the young A-type HR8799 provides a unique laboratory to: a) test planet formation theories;b) probe the diversity of system architectures at these separations, and c) pe...
详细信息
Context. The planetary system discovered around the young A-type HR8799 provides a unique laboratory to: a) test planet formation theories;b) probe the diversity of system architectures at these separations, and c) perform comparative (exo)planetology. Aims. We present and exploit new near-infrared images and integral-field spectra of the four gas giants surrounding HR8799 obtained with SPHERE, the new planet finder instrument at the Very Large Telescope, during the commissioning and science verification phase of the instrument (July-December 2014). With these new data, we contribute to completing the spectral energy distribution (SED) of these bodies in the 1.0-2.5 mu m range. We also provide new astrometric data, in particular for planet e, to further constrain the orbits. Methods. We used the infrared dual-band imager and spectrograph (IRDIS) subsystem to obtain pupil-stabilized, dual-band H2H3 (1.593 mu m, 1.667 mu m), K1K2 (2.110 mu m, 2.251 mu m), and broadband J (1.245 mu m) images of the four planets. IRDIS was operated in parallel with the integral field spectrograph (IFS) of SPHERE to collect low-resolution (R similar to 30), near-infrared (0.94-1.64 mu m) spectra of the two innermost planets HR8799 d and e. The data were reduced with dedicated algorithms, such as the Karhunen-Loeve image projection (KLIP), to reveal the planets. We used the so-called negative planets injection technique to extract their photometry, spectra, and measure their positions. We illustrate the astrometric performance of SPHERE through sample orbital fits compatible with SPHERE and literature data. Results. We demonstrated the ability of SPHERE to detect and characterize planets in this kind of systems, providing spectra and photometry of its components. The spectra improve upon the signal-to-noise ratio of previously obtained data and increase the spectral coverage down to the Y band. In addition, we provide the first detection of planet e in the J band. Astrometric positions fo
暂无评论