Experiments testing the accuracy of patternrecognition techniques require large banks of samples exhibiting the characteristics to be recognized. What is needed is a general-purpose pattern generator. Since totally r...
详细信息
Experiments testing the accuracy of patternrecognition techniques require large banks of samples exhibiting the characteristics to be recognized. What is needed is a general-purpose pattern generator. Since totally random data is of little use when specific classes of patterns are to be recognized the artificial pattern generator must construct samples under user control. The techniques for generating these artificial data are presented. Attention is given to sample representation, types of distortions, and methods for detecting pattern kernels which can be combined to describe gross structure within a prototype. These methods are the key to providing for controlled degrees of distortion while maintaining similarity across the artificial data base.
The goal of feature extraction in patternrecognition is to reduce the dimensionality of the space in which classes of data are represented without greatly reducing the separability of the classes. An approach to feat...
详细信息
The goal of feature extraction in patternrecognition is to reduce the dimensionality of the space in which classes of data are represented without greatly reducing the separability of the classes. An approach to feature extraction based on functions of the class correlation matrices are chosen, the present method extends the methods of feature extraction proposed by K. Fukunaga and W. L. G. Koontz. If certain types of nonlinear functions are employed, the method reduces to the orthogonal subspace method of Wantanabe. Optimization of selected features through selection of appropriate functions is discussed briefly. Preliminary results of classification of radar signatures using the feature extraction methods described here are presented.
Digital imageprocessing has been extended to the enhancement of time-varying picture information. Fluoroscopic medical images recorded on videotape have been processed by digital filtering to enhance cardiac synchron...
详细信息
Digital imageprocessing has been extended to the enhancement of time-varying picture information. Fluoroscopic medical images recorded on videotape have been processed by digital filtering to enhance cardiac synchronous densitometry related to the pulsatile perfusion of blood in lung tissue. The enhancement was obtained by means of a Wiener filter adapted to the power spectrum of the electrocardiogram which is recorded on the audio channel of the videotape. Since the enhancement is of temporal contrast variations as opposed to spatial ones, the resulting information contains valuable physiological data. The authors' experiments have indicated that pulmonary vascular pulsations are diminished in amplitude in human patients with lung carcinoma. These pulsations are not visible to the radiologist without the computer enhancement. The changes in the densiometer signal from vascular pulsations are typically on the order of 0. 5%.
The major problem encountered in scene analysis is that the picture patterns to be examined are always subject to changes in translation, scale, contrast or more generally to so-called ″rubber sheet″ distortions. Th...
详细信息
The major problem encountered in scene analysis is that the picture patterns to be examined are always subject to changes in translation, scale, contrast or more generally to so-called ″rubber sheet″ distortions. Then any scene analysis system using the models that represent and match a picture pattern in the actual image in terms of a fixed template will be infeasible in actual applications. A reasonable solution to the failure of the fixed template model is suggested. A model in the form of a flexible template, in which information concerning the variations that actual visual objects are allowed to have are included in the representation, is used to represent and match the picture patterns. To do this, the template is parameterized in such a way that it will not be sensitive, or even variant, to any of the rubber-sheet distortions. Then by assigning a cost function which corresponds to the way that the set of parameters might be adjusted, a decision scheme is used to evaluate the cost and render the set of optimum parameters that constitute the visual object.
Three standard approaches to automatic texture classification make use of features based on the Fourier power spectrum, on gray level co-occurrences, and on statistics of local properties, respectively. In this paper,...
详细信息
Three standard approaches to automatic texture classification make use of features based on the Fourier power spectrum, on gray level co-occurrences, and on statistics of local properties, respectively. In this paper, features of these types are applied to a set of 54 picture samples taken from aerial photographs of nine terrain types (Lake, Marsh, Orchard, Railroad, Scrub, Suburb, Swamp, Urban, and Woods). Classification results are compared for individual features and pairs of features within each feature class. In general, the Fourier based features performed poorest, while those based on statistics of gray level differences performed best.
Experiences gained while programming several patternrecognition algorithms in the languages ALGOL, FORTRAN, PL/1 and PASCAL are described. The algorithms discussed are for boundary encodings of two-dimensional binary...
详细信息
Experiences gained while programming several patternrecognition algorithms in the languages ALGOL, FORTRAN, PL/1 and PASCAL are described. The algorithms discussed are for boundary encodings of two-dimensional binary pictures, calculating and exploring the minimum spanning tree for a set of points, recognizing dotted curves from a set of planar points and performing a template matching in the presence of severe noise distortions. The lesson seems to be that patternrecognition algorithms require a range of data structuring capabilities for their implementation, in particular arrays, graphs and lists. The languages PL/1 and PASCAL have facilities to accommodate graphs and lists but there are important differences for the programmer. The ease with which the template matching program was written, debugged and modified during a 3 week period, using PASCAL, suggests that this small but powerful language should not be overlooked by those researchers who need a quick, reliable, and efficient implementation of a patternrecognition algorithm requiring graphs, lists and arrays.
A new edge strength evaluation and detection technique for picture analysis is reported. This technique was developed to upgrade performance of the region grower (object locator), based on edge detection. The main mot...
详细信息
A new edge strength evaluation and detection technique for picture analysis is reported. This technique was developed to upgrade performance of the region grower (object locator), based on edge detection. The main motive behind the development of the new edge operation was performance ungrading along the following lines: (1) efficient detection of points that are obviously not edge points; (2) achievement of independence of the edge value from shape and structure of the objects (lines, sharp corners, etc. ); (3) keep the edge evaluation self scaling with respect to linear changes in signal and noise. To achieve those seemingly conflicting goals, a sequential decision approach was used.
An interactive graphics system with bi-directional audio capability has been developed for use in classifying and describing (or characterizing) speech and image data. The use of an interactive system allows human ins...
详细信息
An interactive graphics system with bi-directional audio capability has been developed for use in classifying and describing (or characterizing) speech and image data. The use of an interactive system allows human inspection of intermediate steps in algorithm development, easy modification of parameters, and immediate examination of the results. The graphics portion of the system uses a keyboard, joystick, opaque and transparent image scanners, and control panel for data input and has access to a Tektronix 611 memory scope, an electrostatic printer/plotter, a Calcomp plotter, and a high-resolution CRT with a camera attachment for data output. As many as three different files may be processed concurrently and may be stored either on magnetic tape or disc. A history file which maintains a record of all or selected operations, may be viewed directly or on hardcopy. The system has been used to generate a labelled speech data base of 35,000 tokens for use in semi-automatic speaker identification character recognition algorithm capable of MOS/LSI realization, to analyze imagery for bandwidth reduction studies, and to research classification and matching algorithms for use in fingerprint processing. Examples from each of the application areas are presented.
Digital imageprocessing has been extended to the enhancement of time-varying picture information. Fluoroscopic medical images recorded on videotape have been processed by digital filtering to enhance cardiac synchron...
详细信息
Digital imageprocessing has been extended to the enhancement of time-varying picture information. Fluoroscopic medical images recorded on videotape have been processed by digital filtering to enhance cardiac synchronous densitometry related to the pulsatile perfusion of blood in lung tissue. The enhancement was obtained by means of a Wiener filter adapted to the power spectrum of the electrocardiogram which is recorded on the audio channel of the videotape. Since the enhancement is of temporal contrast variations as opposed to spatial ones, the resulting information contains valuable physiological data. The authors' experiments have indicated that pulmonary vascular pulsations are diminished in amplitude in human patients with lung carcinoma. These pulsations are not visible to the radiologist without the computer enhancement. The changes in the densiometer signal from vascular pulsations are typically on the order of 0. 5%.
Segmentation of a picture into regions of uniform ″visual impression″ is a necessary step in many patternrecognition applications. A data structure is described for efficient segmentation by texture in general, and...
详细信息
Segmentation of a picture into regions of uniform ″visual impression″ is a necessary step in many patternrecognition applications. A data structure is described for efficient segmentation by texture in general, and computing Fourier transforms in particular. This is done by showing ways to estimate the transform over nonrectangular regions in a meaningful way. The basic data structure is formed by dividing the picture recursively into four, sixteen, sixty-four, etc. , parts and then creating a tree whose nodes correspond to regions of the picture. The successors of a node correspond to the four subsets of its region. Such a data structure has already been used for segmentation by average brightness level. This tree is referred to as the picture pyramid. The basis algorithm for segmentation is outlined. In general the use of the split-and-merge procedure in combination with the picture pyramid data structure allows fast updating of the texture descriptors (including the Fourier transform) as the regions are modified during interations of the algorithm.
暂无评论