Showing papers on "Contextual image classification published in 1993"

Pattern Recognition Using Invariants Defined From Higher Order Spectra- One Dimensional Inputs

Abstract: A new approach to pattern recognition using in- variant parameters based on higher order spectra is presented. In particular, invariant parameters derived from the bispec- trum are used to classify one-dimensional shapes. The bispec- trum, which is translation invariant, is integrated along straight lines passing through the origin in bifrequency space. The phase of the integrated bispectrum is shown to be scale and amplifi- cation invariant, as well. A minimal set of these invariants is selected as the feature vector for pattern classification, and a minimum distance classifier using a statistical distance measure is used to classify test patterns. The classification technique is shown to distinguish two similar, but different bolts given their one-dimensional profiles. Pattern recognition using higher or- der spectral invariants is fast, suited for parallel implementa- tion, and has high immunity to additive Gaussian noise. Sim- ulation results show very high classification accuracy, even for low signal-to-noise ratios.

Tropical vegetation mapping with combined visible and SAR spaceborne data

Abstract: Within the tropics, due to high cloud cover, vegetation monitoring can only be achieved by using both visible and microwave spaceborne data In the present study, Shuttle Imaging Radar (SIR-B) data and SPOT data acquired over central Sumatra are used in a supervised classification of vegetation Emphasis is put on SAR data utilization Assessment of SAR data potentialities is carried out by using (1) SIR-B imagery alone, and (2) combined SJR-B and SPOT data The article describes the methodology developed for exploiting the SIR-B image including raw data analysis, image filtering and image classification Finally, combined filtered SIR-B imagery and SPOT data are used jointly in a supervised classification leading to a better discrimination of surface units These results indicate the unique capabilities of combined visible and SAR data for monitoring tropical vegetation, especially for forest/non-forest discrimination

Initial learning of document structure

Abstract: Proposes an approach for automatically generating a decision tree which is applied as a model for the logical labeling of business letters. Instead of top-down determination of the discriminating attributes, the system inspects a finite set of document instances that are presented to a learner in a bottom-up position. The learner itself figures out local similarities, rates them with respect to the overall structure, and determines the best structural match of two instances (neighborhood). The entire decision tree is grown step by step deducing subtrees by forming generalizations from a neighborhood. Consequently, heuristics are learned for structurally discriminating documents during subsequent classification. >

Segmentation and classification for mixed text/image documents using neural network

Abstract: A segmentation and classification method for separating a document image into printed character, handwritten character, photograph, and painted image regions is presented. A document image is segmented into rectangular areas. Each of which contains a cluster of image elements. A layered feed-forward neural network is then used to classify each segmented area using the histograms of gradient vector directions and luminance levels. A high classification performance was obtained, even with a small number of training samples. It is confirmed that the histograms of gradient vector directions and luminance levels are significantly effective features for the classification of the four kinds of image regions. Increasing the number of the discrimination areas improves the classification performance sufficiently even using a small number of training samples for the neural network. >

Moment-preserving edge detection and its application to image data compression

Abstract: In contrast to the numerous edge-detection techniques that detect edges either point by point or using overlapping circular windows, an edge detector using nonoverlapping rectangular windows is proposed. The detector examines the pixels within each rectangular window of an image, and decides whether an edge element is present or not in the window. Based on the gray and mass moment-preserving principles, the step edge is estimated locally to subpixel accuracy using analytical formulas. To apply the edge detection results to image compression, the detected edge elements are then tracked and grouped based on proximity and orientation. Using the line parameters of the grouped edge elements, region boundaries are approximated in a piecewise linear manner. This reduces the amount of data required to describe region shapes and is useful for compressing some types of images. Good experimental results of compressing character and trademark images are also included to show the feasibility of the proposed approach.

Dimensionality reduction by optimal band selection for pixel classification of hyperspectral imagery

Abstract: SUMMARY AND CONCLUSIONSBand selection has been shown here and elsewhere to be a practical method of data reduction for hyperspectral image data.Moreover, band selection has a number of advantages over linear band combining for reducing the dimensionality of highdimensional data. Band selection eliminates the requirement that all bands be measured before data dimensionality isreduced. Bands that are uninformative about pixel classification need not be measured or communicated. Band sets can betailored to specific classification goals (classes, error rates, etc.). Band selection reduces data link requirements, yet retains atunable capability to collect as many bands as required for a specific application. Feature selection algorithms developed forstatistical pattern classifier design can be used to perform band selection. The Greedy (2, 1) feature selection algorithm hasbeen shown to be a practical means of selecting bands. In addition this algorithm has theoretical advantages over the ForwardSequential algorithm, making it the method of choice for hyperspectral applications.

M2dSOMAP: clustering and classification of remotely sensed imagery by combining multiple Kohonen self-organizing maps and associative memory

Abstract: This paper investigates a hybrid neural network framework by combining unsupervised and supervised neural learning paradigms on a unified representation platform of multiple Kohonen 2D self-organizing maps (M2dSOM) with the assistance of associative memory for clustering and classification of remotely sensed (RS) imagery. The M2dSOM is a regional form of such for both cluster region and decision region. A new supervised learning algorithm is proposed that exploits the input portion of supervising samples to discover mismatches between cluster and decision regions by a k-winner selection process and then correct the cluster boundaries based on a majority vote for a new cluster membership from the k winners. Finally, an associative memory is employed to form a mapping between clusters and classification labels by samples. Two association configurations are suggested. Analysis of this mapping SONN model (called M2dSOMAP) in relation to RS imagery analysis with comparison to other methods is briefly discussed.

Fuzzy decision making in the classification of multisource remote sensing data

Abstract: Our objective was to develop a knowledge-based strategy for the classification, considered a cognitive process, of multisource data including remote sensing images. The main feature of our approach is the use of fuzzy sets as the representation framework. This strategy supports two supervised image classification procedures, one based on a fuzzy statistical classifier and the other on a feed-forward fuzzy trained neural network. Approximate reasoning techniques, based on fuzzy production rules, are applied to model the multifactorial evaluation process in which results from the classification of remote-sensing images are integrated with other data. An example of multisource remote-sensing data classification applied in fire prevention is presented together with numerical results and an experimental verification of the approach.

Fuzzy classification of Earth terrain covers using multi-look polarimetric SAR image data

Abstract: The conventional approach of terrain image classification which assigns a specific class for each pixel is inadequate because the area covered by each pixel may embrace more than a single class. Fuzzy set theory which has been developed to deal with imprecise information can provide a more appropriate solution to this problem. In the paper, the authors used the fuzzy c-means clustering algorithm for the segmentation of a polarimetric SAR image. The distance measure utilized in the algorithm was derived from the complex Wishart distribution of the pixel data presented in the covariance matrix format. The algorithm computes the feature covariance matrix for each class and generates a fuzzy partition of the whole image. Classification of the image is achieved using a defuzzification criterion. The results are similar to those of supervised statistical methods. NASA/JPL AIRSAR data is used to substantiate this fuzzy classification algorithm. >

Comparative evaluation of pattern recognition techniques for detection of microcalcifications

Abstract: Computer detection of microcalcifications in mammographic images will likely require a multi-stage algorithm that includes segmentation of possible microcalcifications, pattern recognition techniques to classify the segmented objects, a method to determine if a cluster of calcifications exists, and possibly a method to determine the probability of malignancy. This paper will focus on the classification of segmented objects as being either (1) microcalcifications or (2) non-microcalcifications. Six classifiers (2 Bayesian, 2 dynamic neural networks, a standard backpropagation network, and a K-nearest neighbor) are compared. Methods of segmentation and feature selection are described, although they are not the primary concern of this paper. A database of digitized film mammograms is used for training and testing. Detection accuracy is compared across the six methods.

Combining image classification and image compression using vector quantization

Abstract: The goal is to produce codes where the compressed image incorporates classification information without further signal processing. This technique can provide direct low level classification or an efficient front end to more sophisticated full-frame recognition algorithms. Vector quantization is a natural choice because two of its design components, clustering and tree-structured classification methods, have obvious applications to the pure classification problem as well as to the compression problem. The authors explicitly incorporate a Bayes risk component into the distortion measure used for code design in order to permit a tradeoff of mean squared error with classification error. This method is used to analyze simulated data, identify tumors in computerized tomography lung images, and identify man-made regions in aerial images. >

Pattern classification approach to segmentation of chest radiographs

Abstract: In digital chest radiography, the goal of segmentation is to automatically and reliably identify anatomic regions such as the heart and lungs. Aids to diagnosis such as automated anatomic measurements, methods that enhance display of specific regions, and methods that search for disease processes, all depend on a reliable segmentation method. The goal of this research is to develop a segmentation method based on a pattern classification approach. A set of 17 chest images was used to train each of the classifiers. The trained classifiers were then tested of a different set of 16 chest images. The linear discriminant correctly classified greater than 70%, the k-nearest neighbor correctly classified greater than 70% and the neural network classified greater than 76% of the pixels from the test images. Preliminary results are favorable for this approach. Local features do provide much information, but further improvement is expected when addition information, such as location, can be incorporated.

A symbol recognition system

Abstract: A hierarchical neural network approach is presented for the automatic conversion of image documents (ACID), which specifically describes a prototype symbol recognition system (SRS) for automatic computer processing of electrical engineering drawings. This approach achieves a significant reduction of human involvement in the symbol model encoding and recognition processes in contrast to such traditional approaches based on thinning, line tracing, and other structural feature extraction techniques. A set of image intensity moments, which are invariant to geometric transformations, is used as features. A hierarchical neural classifier demonstrates faster and more accurate capabilities for model encoding and recognition. The test results from hand-drawn images by using templates achieves a recognition rate of 98.5% on training symbols and 89% on test symbols. >

Parallel image classification using multiscale Markov random fields

Abstract: The application of massively parallel multiscale relaxation algorithms to image classification is considered. First, a classical multiscale model applied to supervised image classification is presented. The model consists of a label pyramid and a whole observation field. The potential functions of the coarse grid are derived by simple computations. Then, a scheme which introduces a local interaction between two neighbor grids in the label pyramid is proposed. This is a way to incorporate cliques, with far-apart sites for a reasonable price. Finally, results on noisy synthetic data and on a SPOT image obtained by different relaxation methods using these models are presented. >

Document image segmentation and text area ordering

Abstract: A system for document image segmentation and ordering text areas is described and applied to both Japanese and English complex printed page layouts. There is no need to make any assumption about the shape of blocks, hence the segmentation technique can handle not only skewed images without skew-correction but also documents where column are not rectangular. In this technique, on the bottom-up strategy, the connected components are extracted from the reduced image, and classified according to their local information. The connected components are merged into lines, and lines are merged into areas. Extracted text areas are classified as body, caption, header, and footer. A tree graph of the layout of body texts is made, and we get the order of texts by preorder traversal on the graph. The authors introduce the influence range of each node, a procedure for the title part, and extraction of the white horizontal separator. Making it possible to get good results on various documents. The total system is fast and compact. >

Maximum a posteriori classification of multifrequency, multilook, synthetic aperture radar intensity data

Abstract: We present a maximum a posteriori (MAP) classifier for classifying multifrequency, multilook, single polarization SAR intensity data into regions or ensembles of pixels of homogeneous and similar radar backscatter characteristics. A model for the prior joint distribution of the multifrequency SAR intensity data is combined with a Markov random field for representing the interactions between region labels to obtain an expression for the posterior distribution of the region labels given the multifrequency SAR observations. The maximization of the posterior distribution yields Bayes's optimum region labeling or classification of the SAR data or its MAP estimate. The performance of the MAP classifier is evaluated by using computer-simulated multilook SAR intensity data as a function of the parameters in the classification process. Multilook SAR intensity data are shown to yield higher classification accuracies than one-look SAR complex amplitude data. The MAP classifier is extended to the case in which the radar backscatter from the remotely sensed surface varies within the SAR image because of incidence angle effects. The results obtained illustrate the practicality of the method for combining SAR intensity observations acquired at two different frequencies and for improving classification accuracy of SAR data.

Integration of neural network and statistical image classification for land cover mapping

Abstract: Artificial neural networks and statistical classifiers both give good performance in image classification. Since the two methods are based on significantly different mathematical approaches and have complementary capabilities, a useful solution for optimizing performance is to combine them. A method is presented to integrate both types of classifier. In this method both neural network and maximum-likelihood classifiers are initially trained concurrently with the same data set. A second neural network is then trained using only pixels for which the two classifiers did not initially agree. This second network is thus trained specifically to discriminate ambiguous pixels. In the actual classification a simple procedure is adopted to decide which of the classifiers is best to use for a given pixel. >

Perfect metrics

Abstract: The authors describe an experiment in the construction of perfect metrics for minimum-distance classification of character images. A perfect metric is one that, with high probability, is zero for correct classifications and non-zero for incorrect classifications. They promise excellent reject behavior in addition to good rank ordering. The approach is to infer from the training data faithful but concise representations of the empirical class-conditional distributions. In doing this, the authors have abandoned many visual simplifying assumptions about the distributions, e.g., that they are simply-connected, unimodal, convex, or parametric (e.g., Gaussian). The method requires unusually large and representative training sets, which we provide through pseudorandom generation of training samples using a realistic model of printing and imaging distortions. The authors illustrate the method on a challenging recognition problem: 3755 character classes of machine-print Chinese, in four typefaces, over a range of text sizes. In a test on over three million images, the perfect-metric classifier achieved better than 99% top-choice accuracy. In addition, it is shown that it is superior to a conventional parametric classifier. >

Seafloor characterization using texture

Abstract: Texture analysis is performed on multibeam sonar imagery. A set of 14 texture features is computed using cooccurrence matrices to form the feature space. The dimensionality of the feature space is reduced by extracting the principal components from the original feature space. Classification of the image is performed on the principal components using the K-means algorithm. Results indicate that seafloor bottom types can be characterized by analyzing the texture of bathymetric sonar images. >

European language determination from image

Abstract: The authors have developed a technique for determining the language from an image of text. This work is restricted to a small subset of European languages, but uses techniques which should be applicable across many more languages. The method first makes generalizations about images of characters, then performs gross classification of the isolated characters and agglomerates these class identities into spatially isolated (word) tokens. Analysis of corpora in English, French and German yields training data for a language classifier designed to codify the spatial relationships of the connected components which compose the letter-forms. Linear discriminant analysis provides classification criteria on which the test data are evaluated. The resulting process takes in images of text and produces a language classification based on image representations and generalizations about relative token shape frequency in the target languages. >

Script representation by a generalized skeleton

Abstract: The authors propose a new approach for script skeletonization which exploits the richness of gray level images by the use of the liminosity matrix and its first and second derivatives. They give the general outline of a representation of the information, based on the distinction between regularities and singularities of the writing. This decomposition stems from a multi-criteria stroke following which enables a contextual analysis of the writing. Segmentation between script and background occurs simultaneously with skeletonization. There is no assumption on the nature of strokes and image scanning does not imply preferential orientations. The results depends on sound parameters which have an intrinsic meaning. It provides in addition to representative chained points and their attributes, some symbolic information which enables a contextual interpretation of the script. The extracted data generalizes the skeleton notion. >

Color map image segmentation using optimized nearest neighbor classifiers

Abstract: The author presents an optimized nearest neighbor rule based technique for extracting characters and lines from color geographic map images. In this method, the segmentation procedure is treated as a pattern classification problem. The author first obtains training samples interactively from characters, lines, and the background of an image. One can also produce training samples automatically using clustering algorithms. The author then generates a set of prototypes from the training samples and optimize the prototypes using a multilayer neural network to increase their classification power. The color image is classified pixel by pixel using the optimized prototypes. The method has been compared with adaptive thresholding with favorable results. >

Classification of high dimensional multispectral image data

Abstract: A method for classifying high dimensional remote sensing data is described. The technique uses a radiometric adjustment to allow a human operator to identify and label training pixels by visually comparing the remotely sensed spectra to laboratory reflectance spectra. Training pixels for material without obvious spectral features are identified by traditional means. Features which are effective for discriminating between the classes are then derived from the original radiance data and used to classify the scene. This technique is applied to Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) data taken over Cuprite, Nevada in 1992, and the results are compared to an existing geologic map. This technique performed well even with noisy data and the fact that some of the materials in the scene lack absorption features. No adjustment for the atmosphere or other scene variables was made to the data classified. While the experimental results compare favorably with an existing geologic map, the primary purpose of this research was to demonstrate the classification method, as compared to the geology of the Cuprite scene.

Machine vision techniques for subpixel estimation of critical dimensions

Abstract: Novel edge detection and line-fitting machine vision algorithms are applied for linewidth measurement on optical images of integrated circuits. The techniques are used to achieve subpixel resolution. The strategy employs a two-step procedure. In the first step, a neural network is used for edge detection ofthe image. Three neural network approaches are investigated: self-organizing, bootstrap linear threshold, and constrained maximization strategies. The weights of the neural networks are estimated using unsupervised learning procedures, the advantage of which is the ability to adapt to the imaging environment. Consequently, these proposed neural network approaches for edge detection do not require an a priori data base of images with known linewidths for calibration. In the second step, line-fitting methods are applied to the edge maps defined by the neural network to compute linewidth. Two methods are investigated: the Hough transform method and an eigenvector strategy. By employing this two-step strategy, the entire image is used to estimate linewidth as opposed to the use of just a single or a few line scans. Thus, edge roughness effects can be spatially averaged to obtain an optimal estimate of linewidth, and subpixel resolution can be achieved. However, the accuracy (or variance) of this estimate will, of course, be dependent on issues such as pixel size and the capability of the imaging system. The techniques are general and can be used on images from a variety of microscopes, including optical and electron-beam microscopes.

Neural clustering for optimal KLT image compression

Abstract: A multiple class approach is proposed for improving the performance of the Karhunen-Loeve transform (KLT) image compression technique. The classification is adaptively performed by suitable neural networks. Several examples are presented in order to show that the proposed method performs much better than the classical discrete cosine transform (DCT). >

Mathematical morphology, granulometries, and texture perception

Abstract: In this work we present a classical morphological tool, granulometry, and a practical application on medical images, pneumoconiosis classification. The radiologist diagnose on these images is based on a preattentive discrimination process of the textural patterns appearing at the pulmonar parenchyma. Thus, in order to automatize this classification we have chosen a tool which agrees with perceptual theories of Computer Vision on texture discrimination. Our work is centered, concretely, on the perceptual models based on texton theory. These works base texture discrimination on differences in density of texton attributes. We link this approach with a morphological tool, granulometry, as a helpful multi-scale analysis of image particles. The granulometric measure provides a density function of a given feature, which depends on the family of algebraic openings selected. Thus in this paper we defined different granulometries which allow us to measure the main texton features, such as, shape, size, orientation or contrast, proposing a granulometric analysis as a systematic tool for texture discrimination according to a perceptual theory. And finally, we present the application of measuring size density on some radiographic images suffering from pneumoconiosis.© (1993) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.