In this paper, a classification method for classifying subjects’ ability to follow some predefined trajectories is proposed
using phase only filter (POF) . In this research, we use three predefined trajectory patterns of different difficulty levels,
and a set of data comprising four different classes of movements. We propose a POF to classify the data in those classes.
POF can be assumed as a Complex Match Filter (CMF) where the amplitude of the object function is set to unity. The POF
searches the entire image to find a match to the input filter. The trajectory in all three patterns contains edges and sharp
turns which could considerably help to distinguish between the classes. Therefore, in this method, the reference pattern is
segmented to several parts and critical segments of the trajectory used as an input filter or the pattern to search for. The
classification task is applied for each pattern separately and the results obtained are fused based on different weights. The
optimum weights for the fusion are obtained by using the training data and the linear regression technique.
A feature extraction-based classification method is proposed in this paper for verifying the capability of human’s neck in target tracking. Here, the target moves in predefined trajectory patterns in three difficulty levels. Dataset used for each pattern is obtained from two groups of people, one with whiplash associated disorder (WAD) and asymptomatic group, who behave in both sincere and feign manner. The aim is to verify the WAD group from asymptomatic one and also to discriminate the sincere behavior from the feigned one. Sparse and low-rank feature extraction is proposed to extract the most informative feature from training samples and then each sample is classified into the group which has the highest correlation coefficient with. The classification results are improved by fusing the results of the three patterns.
This paper examines the shadow extraction problem associated with satellite images; namely the fact that
images taken at a different time include different shadow components. In the present research we attempt to define the
shadow component of an image in a controlled lab environment in terms of the phase of the Fourier spectrum of the
image, devise a novel method for extracting the simple shadow component from the image, and to create a meta-image
suitable for change detection post processing.
This paper examines the shadow extraction problem associated with satellite images; namely the shadow component
of the images. The present research extends our previous work on shadow extraction in simple scenes. Extraction of
multivariable shadows caused by earth's rotation and satellite viewing angle is examined. Attempt is made to define the
variables of such shadow component in a controlled lab environment in terms of the phase of the Fourier spectrum of the
image, device a novel method for extracting complex shadow components from the image, and to create a meta-image
suitable for change detection post processing.
Today’s sensor networks provide a wide variety of application domain for high-speed pattern classification systems. Such high-speed systems can be achieved by the use of optical implementation of specialized POF correlator. In this research we discuss the modeling and simulation of the phase only filter (POF) in the task of pattern classification of multi-dimensional data.
The design and development of binarized composite binary phase only filters for three-dimensional image identification is discussed in this paper. A composite binary phase only filter based algorithm is developed to detect an unknown image from the pre-constructed database. The angular position of the unknown image is found by a cross correlation technique. The formulation of composite filters reduces the number of correlation operations necessary for detection.
Automatic Target Tracking continuous to be an issue of major interest in the defense industry. However, evaluation of scenes and quantification of algorithm performance on different scenes continuous to be a challenging task. In this research we have developed a platform, or test bed, to test and compare the performance of various algorithms. Using this platform, different algorithms can be compared. Fifty unclassified scenes of different complexity provided by the Army Research Office are analyzed using this platform and phase only filter (POF) combined with two different functions defining the Region Of Interest (ROI). The ROI functions significantly speed up the search process.
This article presents an application of Phase Only Filter (POF) to the classification of volatile compound samples with chemical sensor arrays. Sinusoidal temperature modulator excites the chemical sensor array. A system composed of multiple sensors for data acquisition requires the analysis of multiple data signals in order to classify the input data. One such system is the Electronic Nose system (eNose). The eNose data is in fact 1D data from multiple data sources. In this work five samples of three different kinds of coffee are used to build an odor database. An unknown test sample is then classified. In this research data, from such system will be analyzed and classified using the POF and compared to the more conventional K Nearest Neighbors (KNN) classifier. Both classifiers use only the mean and standard deviation as the classification feature space. The difference of the correlation between odor database and the auto correlation of the test sample is the measure of closeness for the POF. For the KNN, the measure of closeness is the difference between the odor database and the test sample.
This paper investigates the role of composite filters in reducing the search time for 3D model based object recognition. When one moves from 2D to 3D, one is faced with a huge amount of information to deal with. The composite filter combined with a multistage scheme is developed for processing this huge information. The design scheme for the composite filters is also elaborated. The procedures discussed in this paper demonstrate how detection of these various model images might help formulate a new metric for recognition performance.
In this paper, a novel metric is defined that will allow one to compare the performance of 3-D pattern recognition systems. Any real object is inherently, three-dimensional. Therefore, any input object for an automated target recognition system should be ideally compared to the 3-D information about the object. The proposed metric captures the essence of such comparisons.
A novel uni-complex valued trinary associative model, which is implementable in the optical domain, is proposed. Retrieval of the stored pattern is accomplished using an threshold formula in the inner product domain. An algorithm to determine adaptive threshold formula for this trinary associative memory model is presented. The optimal threshold is chosen to yield the best performance. Different threshold parameters have been investigated to obtain the range of optimal threshold parameters. In order to validate our performance model, character recognition problem with noisy and noise-free data are investigated. Moreover, a bi-complex representation model for associative memory retrieval is presented and compared to previous methods.