This paper presents a cooperative hierarchical fusion scheme based on `d trous' wavelet transformation for the fusion of infrared and visible images. At first, the restoration algorithm of multi-frames is presented to filter image noise and to improve the detail information of the infrared image. Then both the infrared and the visible images are decomposed to multilayer wavelet planes, respectively. A hierarchical merging is used for feature selection of wavelet planes by taking the weighted average at each scale. Lastly, the inverse wavelet transformation is implemented from the approximation data of the infrared image and the fused wavelet coefficients at various scales. One visible and three-frame infrared images are used to test the performance of the proposed scheme. Experimental results show that the spatial resolution improvement of the infrared image can be cooperatively achieved by multi-frame image restoration and sensor fusion. The advantage of the proposed cooperative merging algorithm is that the salient detail information from both visible and infrared images is preserved. No artifacts such as the blocking effect exist in the merged result. Moreover, the proposed method allows use of a dyadic wavelet to merge different sensor data of nondyadic resolution in a simple and efficient approach.
The problem of registering two successive video frames has traditionally been addressed with the 2D cross-correlation shift estimator. In this paper, a computationally efficient method of image registration is investigated that can achieve improved registration performance over the 2D cross- correlator in the presence of both fixed-pattern and temporal noise. This is accomplished by transforming both the current frame and the previous frame into two vector projections formed by accumulating pixel values across the rows and the columns of the array. The 2D cross-correlator requires three 2D Fourier transforms at the size of the image. In order to avoid the use of 2D Fourier transforms for large arrays, other shift estimation procedures have been developed that rely only on gradients between the two frames to infer the inter-frame shifts. Gradient-based techniques exhibit degraded performance in comparison to the 2D cross-correlator since the gradient operation amplifies noise. The projection-based estimator alleviates the computational burden of estimating shifts while improving the performance relative to the 2D cross-correlation shift estimator. In order to demonstrate the noise rejection capability of the projection-based shift estimator, a figure of merit is developed that reflects the signal-to-noise ratio for the two different shift estimation procedures. The relative performance of the 2D cross-correlation shift estimator and the projection-based shift estimator can be compared through their associated figures of merit. These two methods are also compared through computer simulation.
This paper presents results on appearance based 3D object recognition accomplished using Independent Component Analysis (ICA). A database of images captured by a ccd camera was used. The workspace was then sampled in a certain manner. Features were extracted from the sampled image using ICA employing information maximization approach reported recently. The features of all the objects thus obtained were saved in a database which formed the workspace manifold. The test images was also represented in a similar manner. Recognition was then performed by locating the closest point in the manifold using radial basis function network, which gave the identity and view (or pose) of the object. The use of ICA, in place of principle component analysis is expected to give a `natural' manifold with maximum significant information with least redundancy.
Many correlation filters have been designed to be invariant to certain parameters within a set of input images. For example, the construction of rotation and scale invariant filters is well documented. However, an estimation of a generic varying parameter is not available by these methods. This paper presents a two-filter method that estimates the value of a varying parameter in the input image.
Texture classification is a deceptively simple problem which allows for a multitude of approaches. Many of these methods often are computationally intensive, since they rely on multiscale filtering to develop a feature vector that can be used as an input to a classifier. In this paper, a new feature vector derived from the level sets of a texture is used to discriminate between textures. The new approach is very simple computationally, and provides excellent results. In addition, use of level sets results in a feature which is rotationally invariant.
This paper presents an approach to textured image segmentation based on a combination of fractal features. The traditionally fractal dimension obtained from the original and pre-processed images and is combined with fractal transformation coefficients. The latter are extracted using methodologies from image compression using Iterated Function Systems (IFS). Experiments have shown that coefficients used in IFS for image reconstruction can also be used as segmentation and classification features. Experimental results with various natural texture measures will be presented.
Ordinal optimization is a relatively new field of mathematics that seems never to have been applied to optics. Optics has made extensive use of traditional cardinal optimization. This paper explores the possibility that ordinal optimization might be useful in optics. The conclusion is: Not directly but indirectly.
In this paper, we propose a new directional 3D interpolation algorithm for brain magnetic resonance images. Typically, brain images consist of a number of 2D images. In order to reconstruct 3D objects from slices of 2D images, interpolation operation is required. In most interpolation algorithms in the 3D space, the interpolation operation is performed separately in each coordinate that is orthogonal to each other. However, since the shape of the brain is roughly a sphere, interpolation along these three orthogonal coordinates may result in some information loss, particularly when gradients of pixel values have directions similar to the directions of the coordinates. In order to address this problem, we propose a new directional interpolation algorithm. In the proposed algorithm, we first perform the interpolation along two orthogonal coordinates. Typically, the two orthogonal coordinates would be the coordinates of the 2D images. And then, in order to find the best interpolation of the remaining coordinate, we search various directions that are not orthogonal to the two orthogonal coordinates using cost functions. Experiments show promising results.
This is a paper about a paradigm for invention used by me but also much earlier by the giants of physics and mathematics. The presentation is simple: announce the theme, illustrate it with well known examples, and then show how to use it. Strangely enough, the general scheme has had no name or attention given to it despite its many examples.
It has been a desire of a design engineer to combine various tools of analysis and apply them on one problem at hand. In this paper, we propose an algorithm that combines two signal processing analysis tools: higher-order spectral analysis and wavelets. The computation of polyspectra using conventional approaches involves the use of FFT algorithm. It has been shown that discrete Fourier transform (DFT) can be implemented by a fast algorithm using wavelets. By using this algorithm, polyspectra computational complexity for a certain class of signals reduces to lesser number of computations. In actual implementation, the wavelets in use have to be carefully chosen to balance the benefit of pruning of insignificant data and the price of the transform. Clearly, the optimal choice depends on the class of the data we would encounter. In this paper, we first present an introduction f higher-order spectral analysis. Then we discuss wavelet-based fast implementation of DFT and its importance from higher-order spectral analysis viewpoint. Finally, we develop wavelet-based algorithm for computational of polyspectra followed by conclusions.
We study the solution of block system Tm,nx equals b by preconditioned conjugate gradient methods where Tm,n is an m X m block Toeplitz matrix with n X n Toeplitz blocks. This kind of systems occur in a variety of applications, such as the 2D digital signal processing and the discretization of 2D partial differential equations. We propose a new preconditioner for this kind of block systems. Our preconditioner is defined as the sum of block Toeplitz matrix with Toeplitz blocks and three sparse matrices with structure. Our numerical tests show that our preconditioner is superior to Level-1 and Level-2 circulant preconditioners.
The multi-channel image processing system on the Space Solar Telescope (SST) is described in this paper. This system is main part of science data unit (SDU), which is designed for dealing with the science data from every payload on the SST. First every payload on the SST and its scientific objective are introduced. They are main optic telescope, four soft X- ray telescopes, an H-alpha and white light (full disc) telescope, a coronagraph, a wide band X-ray and Gamma-ray spectrometer, and a solar and interplanetary radio spectrometer. Then the structure of SDU is presented. In this part, we discuss the hardware and software structure of SDU, which is designed for multi-payload. The science data scream of every payload is summarized, too. Solar magnetic and velocity field processing that occupies more than 90% of the data processing of SDU is discussed, which includes polarizing unit, image receiver and image adding unit. Last the plan of image data compression and mass memory that is designed for science data storage are presented.
The work described in this paper addresses the problem for extracting bispectrum feature of speech data. Very often the bispectrum feature extraction and data reduction are complicated due to some limiting constraints, i.e., no prior knowledge of feature's distribution and higher dimensionality of bispectrum data. In this article we developed an adaptive feature extraction mechanism based on cascade neural network in conjunction with feature's dimensionality reduction based on Karhunen-Loeve transformation technique. An adaptive codebook generation algorithm which is a cascade configuration of SOFM (Self Organizing Feature Map) and LVQ (Learning Vector Quantization) was used before the K-L transformation. The transformation was experimentally shown as an effective procedure for orthogonalization and dimensionality reduction of spectrum feature. Performance of our speaker identification system was perceived to be significantly increased even though using limited number of channels in noisy environment. We also tried to improve the capability of adaptive codebook generation algorithm by applying simplified differential competitive learning network.
The wavelet transform theory emerged in recent years has been applied in many fields, such as image compression, edge and feature detection, image enhancement and texture analysis. We have studied some of the developments that lead to the current state of image enhancement, sufficiently considering the configurations of video image and eyespot requirement. This paper presents a new method to accentuate image using multi-scale analysis, time-frequency wavelet transform and minimum distance criteria. After decomposing an image into components of different size, position, and orientation, the amplitude of coefficients in the wavelet transform domain can be altered prior to obtaining the inverse transform. This can selectively accentuate interesting components at the expense of undesirable ones. Experimental results have shown that simpler algorithm, faster running speed, rather improved SNR and contrast can be obtained with this method, resulting in a good enhancement effect.
In this paper, we introduce a linear statistical model for an multispectral image segmentation to simultaneously describe not only interband but also interclass properties. Since the number of parameters to be estimated is reduced compared with the conventional maximum likelihood classifier based on multidimensional normal distributions, it is expected that reliable classification results will be achievable even from a restricted number of training data. We demonstrate the effectiveness of our classifier by applying it to simulated and actual satellite data.
This paper describes an Opto-Silicon Adaptive Imaging (OSAI) system capable of operating at low light intensities with high resolution, high accuracy, wide dynamic range, and high speed. The system consists of three major subsystems: (1) an adaptive imaging system in which a liquid crystal wavefront corrector measures image quality based on statistical analysis of a speckle field; (2) an image quality analyzer (IQA); (3) an opto-silicon multi-chip module combining a high-resolution ferroelectric liquid crystal SLM, CCD photodetector array, field-programmable gate array, and digital signal processor. The OSAI wavefront control applies adaptive optoelectronic feedback for iterative wavefront restoration and distortion compensation, suing an image quality metric based on statistical properties of the speckle field produced by moving a diffuser in the Fourier transform plane of a IQA optical system. A prototype IQA system was designed, manufactured, and tested using an input liquid crystal SLM, a Fourier lens, a light-shaping diffuser, and an output photodiode.
The physical objective is to create, for a time span of about an hour, an antenna reflector with an equivalent aperture diameter of about 300 kilometers, as shown in the last part of the introduction. To do this with the convex ocean surface requires use of ocean surface waves. The oweda system determines the surface wave source for a given reflectivity (from a floor slide to ship wakes). The most useful part of the oweda system is coordination of highly data intensive post processing issues involved with transient structural analysis. Transient analysis automation tends to hide details and provide too much data. OWEDA performs calculations and accumulate statistics during the analysis of the data so that extra storage is not necessary. The owedaState.c report takes far less space than one time step of data, S(n) equals (Theta) (n) and the operations run in tandem, T(n) equals (Theta) (n). The time complexity analysis shows how loose bounds on some processes can be overcome by tight bounds on others for an overall tight time complexity bound on the algorithm.