A two-region segmentation method is described which separates the object at a given location from the surrounding background. To accomplish segmentation, statistics are collected from the background relative to gray l...
详细信息
A two-region segmentation method is described which separates the object at a given location from the surrounding background. To accomplish segmentation, statistics are collected from the background relative to gray level, edges, and texture. Points in the object not matching the background are declared object points. Results are shown for 16 test images. A method of classification of the segmented objects using projections is next presented and discussed.
A system for real-time object recognition and learning from known and unknown classes is presented. A three-level algorithm is described. Level I does not classify patterns. Using deterministic rules, it acts as a pre...
详细信息
A system for real-time object recognition and learning from known and unknown classes is presented. A three-level algorithm is described. Level I does not classify patterns. Using deterministic rules, it acts as a prefilter to eliminate obvious classes from further consideration. Level II algorithm is a modified version of the (k,k prime ) nearest neighbor algorithm designed to deal with objects that do not belong to any of the known classes. Level III algorithm is clustering-oriented. When a new pattern arrives, a clustering algorithm is activated to decide whether to assign this pattern to one of the known classes or to classify it as a new type. Experimental results with the system are reported.
The rotational axes of symmetry of an object can be determined automatically by analyzing which orientations of the object have the same appearance. Knowledge of rotational symmetry is important for classifying object...
详细信息
The rotational axes of symmetry of an object can be determined automatically by analyzing which orientations of the object have the same appearance. Knowledge of rotational symmetry is important for classifying object orientation. An axis of symmetry and its repetition angle can be proposed using one pair of orientations with equal appearance. Proposed axes can be confirmed by verifying predicted equivalent appearances for all sampled orientations. Unless it can be explained by symmetry, the occurrence of equal appearances for two orientations would indicate that the resolution of appearance features is inadequate for the orientation classification task.
Over the last several years a classifier for earth observational image data has been under development which is intended to achieve improved performance by utilizing spatial characteristics of the data as an adjunct t...
详细信息
Over the last several years a classifier for earth observational image data has been under development which is intended to achieve improved performance by utilizing spatial characteristics of the data as an adjunct to multispectral ones. An overview is presented of the conception, development, evaluation, and documentation of this spectral/spatial classifier. The research program leading to this classifier is described, the algorithms of the current implementation called ECHO are outlined, and the results on its performance are summarized. These results show it to have improved accuracy, with greater computation efficiency, and only slightly increased operator complexity.
A natural way of describing pictures made up of line-like curvilinear elements is outlined. A square (or rectangular) rectangular0 frame is assumed to contain the picture. Canvas and grain sizes are specified as limit...
详细信息
A natural way of describing pictures made up of line-like curvilinear elements is outlined. A square (or rectangular) rectangular0 frame is assumed to contain the picture. Canvas and grain sizes are specified as limits for growth or subdivision of the frame respectively. A set of partitioning functions are utilized to extract appropriate regions of the frame with respect to which relationships among primitives or picture fragments are depicted. The complete picture is described in terms of picture fragments and picture fragments in terms of primitives. The picture description thus generated is suited for generation and recognition of pictures.
An approach is presented which provides both efficient storage and complexity characteristics for syntactic patternrecognition problems which can be represented by space-efficient one-dimensional contours. In the cla...
详细信息
An approach is presented which provides both efficient storage and complexity characteristics for syntactic patternrecognition problems which can be represented by space-efficient one-dimensional contours. In the classical k-neighbor approach, the classification of an unknown sample is determined by the already specified labels of the samples of its k-nearest neighbors, where nearness is defined by a Euclidean distance metric. In the proposed approach, however, the nearness criterion used is a similarity measure constructed between sets of samples (primitives) and relations formed among the given primitive and the k most similar primitives to it, where sets of primitives will be similar if they together form a higher level pattern. An example utilizing this approach is described and possible extensions explored. The relevancy of this approach to applications in syntactic scene analysis is also discussed.
An image search task is to determine the locations where a given feature occurs in a digital image, if it occurs at all. A brief review of correlation-based template-matching techniques and of image registration techn...
详细信息
An image search task is to determine the locations where a given feature occurs in a digital image, if it occurs at all. A brief review of correlation-based template-matching techniques and of image registration techniques using pyramids is given. A class of hierarchical search operators which work in pyramids is described and several of them are compared. In particular, policies for setting thresholds, policies for recursive refining of location information, and choices of initial and terminal resolution levels for search are compared. The results suggest that a rich class of search operators exists and should be exploited for image analysis applications.
A series of studies are being performed with the following objectives: (1) assist the physician in the extraction of additional information from laboratory data;(2) present laboratory data to physician in a more conci...
详细信息
A series of studies are being performed with the following objectives: (1) assist the physician in the extraction of additional information from laboratory data;(2) present laboratory data to physician in a more concise, readily interpretable form;and (3) quantitate the utility of various laboratory tests in specific diagnostic situations. Application of several patternrecognition techniques to the problem of microorganism identification from antibiotic sensitivity results showed that classification accuracy could be significantly reduced if the organism were resistant to multiple antibiotics. In clinical chemistry, a data presentation technique based on optimal discriminant vector projections appears promising.
Large English texts on ten differnt subject matters were compiled. Estimates were obtained of the n-gram probability distributions, the word-length for each of the texts as well as English as a whole. Experiments were...
详细信息
Large English texts on ten differnt subject matters were compiled. Estimates were obtained of the n-gram probability distributions, the word-length for each of the texts as well as English as a whole. Experiments were done to test for pairwise differences of the ten texts. Principal component analysis and hierarchical clustering analysis were applied to the data in order to discover any possible similarities and dissimilarities among the different texts. Estimates were obtained of first, second, and third-order entropies for each text, and the texts were tested for pairwise differences according to their first-order entropy estimates. The results are of interest to researchers in psychology, biology, anthropology, and computational linguistics as well as patternrecognition.
Some properties of the separable filter resulting from successive applications of a one-dimensional median filter on the rows and columns of an image are investigated. Although the output of this separable filter is n...
详细信息
Some properties of the separable filter resulting from successive applications of a one-dimensional median filter on the rows and columns of an image are investigated. Although the output of this separable filter is not identical to the corresponding two-dimensional median filter with a square window, its performance in image noise smoothing is very close. In particular, its effectiveness in smoothing spiky noise and behavior with edges is compared with that of the two-dimensional median filter. Besides being more efficient in a computer simulation, the separable filter has a much simpler implementation in real time hardware (at video rates for example), than the corresponding two-dimensional median filter.
暂无评论