Showing papers on "Contextual image classification published in 1990"

Fuzzy supervised classification of remote sensing images

Abstract: A fuzzy supervised classification method in which geographical information is represented as fuzzy sets is described The algorithm consists of two major steps: the estimate of fuzzy parameters from fuzzy training data, and a fuzzy partition of spectral space Partial membership of pixels allows component cover classes of mixed pixels to be identified and more accurate statistical parameters to be generated, resulting in a higher classification accuracy Results of classifying a Landsat MSS image are presented, and their accuracy is analyzed >

Fast algorithms for texture analysis using co-occurrence matrices

Abstract: Texture analysis may be of great importance for the problem of image classification and recognition. Co-occurrence matrices are quite effective for discriminating different textures but have the disadvantage of a high computational cost. In the paper a fast algorithm for calculating parameters of co-occurrence matrices is presented. This method has been applied to the problem of classification and segmentation of artificial and natural scenes: the classification, based on co-occurrence matrix parameters, is implemented pixel-by-pixel by using supervised learning and maximum likelihood estimates. The problem of texture boundary recognition has also been considered and a classification scheme based on more than one window for each pixel is presented. Experimental results show the improvements of classification rates that can be achieved by using this method when compared to a single-window classification.

Fuzzy sets-based classification of electron microscopy images of biological macromolecules with an application to ribosomal particles.

Abstract: SUMMARY Pattern recognition methods based on the theory of fuzzy sets are tested for their ability to classify electron microscopy images of biological specimens. The concept of fuzzy sets was chosen for its ability to represent classes of objects that are vaguely described from the measured data. A number of partitional clustering algorithms and an extensive set of cluster-validity functionals (some already reported and some newly developed) have been applied to a test-data set and to two real-data sets of images. One of the real-data sets corresponded to images of the Escherichia coli 50S ribosomal subunits depleted of proteins L7/L12 and the other set to images of the E. coli 70S monosome in the range of overlap views. These two latter sets had been previously studied by another clustering methodology. The new results obtained by the application of fuzzy clustering techniques will be compared to those previously obtained and some conclusions about the consistency of these classifications will be drawn from this comparison.

Markov random field texture models for classification

Abstract: Two novel approaches to texture classification based upon stochastic modeling using Markov Random Fields are presented and contrasted. The first approach uses a clique-based probabilistic neighborhood structure and Gibbs distribution to derive the quasi likelihood estimates of the model coefficients. Likelihood ratio tests formed by the quasi-likelihood functions of pairs of textures are evaluated in the decision strategy to classify texture samples. The second approach uses a least squares prediction error model and error signature analysis to model and classify textures. The distribution of the errors is the information used in the decision algorithm which employs K-nearest neighbors techniques. A new statistic and complexity measure are introduced called the Knearest neighbor statistic (KNS) and complexity (KNC) which measure the overlap in K-nearest neighbor conditional distributions. Parameter vectors for each model, neighborhood size and structure, performance of the maximum likelihood and K-nearest neighbor decision strategies are presented and interesting results discussed. Results from classifying real video pictures of six cloth textures are presented and analyzed.© (1990) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Problems and approaches for tissue segmentation in 3-D MR imaging

Abstract: The paper describes applications problems and approaches for image segmentation in magnetic resonance imaging. The methods which are proposed work on 3D-datasets with the goal of isolating tissue volumes. Unlike 2D-techniques which operate on multispectral image data in 3D-image segmentation only one image for one anatomical slice is available lacking essential information for tissue discrimination. Three different approaches for the task of volume segmentation are presented. The first is based on the detection of edge structures dividing the original images into anatomically relevant object regions. A 3D-region merging algorithm is applied to extract those regions which belong to the object to be segmented from the region dataset. The second method consists of a polynomial classification of imagepixels into several user-defined tissue classes. Local texture properties are used as discrimination features. The third approach region classification may be regarded as a combination between edge detection and pixel classification. On the basis of a presegmentation of the dataset into object regions a classification process tries to group the regions into different object classes making use of various region features. The latter strategy has yielded the best results and highest reliability for 3D-image segmentation. Further improvements towards minimization of user-interaction are proposed. 1.© (1990) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Translation, rotation and scaling invariant object and texture classification using polyspectra

Abstract: The problem addressed in this paper is the detection and classification of deterministic objects and random textures in a noisy scene. An energy detector is developed in the cumulant domain, by exploiting the noise insensitivity of higher-order statistics. An efficient implementation of this detector is described, using matched filtering. Its performance is analyzed using asymptotic distributions in a binary hypothesis testing framework. Object and texture classifiers are derived using higher-order statistics. They are minimum distance classifiers in the cumulant domain, and can be efficiently implemented using a bank of matched filters. Further, they are robust to additive Gaussian noise and insensitive to object shifts. Extensions, which can handle object rotation and scaling are also discussed. An alternate texture classifier is derived from an ML viewpoint, that is more efficient at the expense of complexity. The application of these algorithms to texture modeling is shown and consistent parameter estimators are obtained. Simulations are shown for both the object and the texture classification problems.© (1990) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Soil cover determination using color image analysis on skidding cost

Abstract: Image classification algorithms using color information were developed to estimate the percent soil cover. The color information was obtained using band-pass optical filters with a black-and-white imaging system. The optimum wavelength pair which best characterizes the color difference between soil and crop canopy was 550 nm and 700 nm. Two classification algorithms, one with the linear discrimination method and the other with the nearest neighbor method, were compared with a manual, photographic grid method and a textural image classification method. Both algorithms measured percent soil cover quickly and accurately. The linear discrimination algorithm performed best in all aspects provided that the image is linearly separable in terms of color. The textural classification method is recommended when a color difference is not evident..

Multispectral tissue classification of MR images using sensor fusion approaches

Abstract: Multi-spectral image data fusion techniques for tissue classification of magnetic resonance (MR) images are presented. Using MR it is possible to obtain imagesof proton density the spin-lattice relaxation time constant ( T1) and the spin-spin relaxation time constant (T2) of the same anatomical section of the human body. In this paper we adopt a sensor fusion approach to tissue classification and segmentation in which each of the three images are treated as the output of different sensors. Regions of the images are modeled as noncausal Gaussian Markov random fields (GMRFs) and the underlying tissue label image is also assumed to follow a Gibbs distribution. Two different multi-spectral tissue labeling algorithms maximum a posteriori (MAP) estimation and the Dempster-Shafer evidential reasoning technique are presented. In the Bayesian MAP approach we use an independent opinion pool for data fusion and a deterministic relaxation to obtain the MAP solution. In practice the Bayesian approach may be too restrictive and a likelihood represented by a point probability value is usually an overstatement of what is actually known. In the Dempster-Shafer approach we adopt Dempster''s rule of combination for data fusion using belief intervals and ignorance to represent our confidence in a particular labeling and we present a new deterministic relaxation scheme that updates the belief intervals. Results obtained from real MR images are presented. 1.© (1990) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

A two-dimensional shift invariant image classification neural network which overcomes the stability/plasticity dilemma

Abstract: A neural network for two-dimensional visual pattern learning and classification is outlined. The new architecture combines the important aspects of two previously developed network designs, simultaneously taking advantage of the unique properties of both. The structure of the Neocognitron network is incorporated to allow shift-invariant and partial scale-invariant recognition, while the top-down attentional and matching mechanisms found in the adaptive resonance theory (ART) model are used to solve the stability-plasticity dilemma. The new network is self-organizing, shift invariant and able to switch automatically between its stable and plastic modes. Computer simulation results for a group of edge extracted patterns are detailed. The neural design uses viable neural mechanisms similar to those thought to exist in biological neural systems

A Spectral-textural Classifier For Digital Imagery

Abstract: Spectral classification is commonly used in remote sensing as a means of extracting information from an image. Unfortunately, the desired classes cannot always be defined by their spectral properties. To overcome this problem, many texture classification methods have been developed. Nonetheless, no single method is widely accepted in remote sensing: some methods do not work with multiband data, some cannot extract irregularly-shaped, textured areas, some do not consider spectral properties, and some are too specific, resulting in scene dependency. In this study, a new classification method is developed, the spectral-texture classifier. The spectral-texture classifier works with multiband images, uses both spectral and spatial characteristics of the class, performs area-based rather than pixel-based classifications, and classifies irregularly-shaped areas. Landsat's Thematic Mapper data were used to test the spectral-texture classifier against the maximum likelihood classifier. It was found that the overall accuracy of classifying land cover with the spectral-texture classifier was higher than that with the maximum likelihood method. Moreover, since the spectral-texture classifier is an area-based classifier, its results do not exhibit the unrealistic isolated spots which degrade the results from single-pixel classifiers such as the maximum likelihood. Since the spectral-texture classifier considers both the spectral and the spatial characteristics of neighboring pixels simultaneously, the classification of textured areas, including the irregularly shaped textured areas, is accurate.

Cluster analysis of medical magnetic-resonance images data: diagnostic application and evaluation

Abstract: We describe the application of statistical clustering algorithms (approximate fuzzy C-means (AFCM) and ISODATA) and a Bayesian/maximum likelihood (BfML) classifier for data dimension reduction and information extraction with MRI. Analyses were performed on 140 cranial and 6 body MR image data sets obtained at 1.5 Tesla (GE Signa) with a variety of pathologies. Cluster analysis methods were run in an unsupervised mode and used to segment image data sets into 32 classes. Unsupervised classification of new image data sets was achieved by training the B/ML classifier on the 32 cluster data set and using the second-order statistics to assign each new image pixel to a cluster centroid in feature space. A translation table was then used to combine these cluster assignments into nine "superclusters" or tissue types. Tissue classification results were evaluated using visual assessment by a radiologic expert and by statistical comparison with a "gold standard" tissue map. Comparison of the newly classified data to the gold standard image using a confusion matrix showed an overall accuracy of 91%. We have found that this approach can improve the diagnostic specificity of MRI and can be applied to new data in an unsupervised mode with a high degree of accuracy.

Adaptive search for morphological feature detectors

Abstract: A closed-loop hybrid learning system that facilitates the automatic design of a multi-class pattern recognition system is described. The design process has three phases: feature detector generation feature set selection and classification. In the first phase a large population of feature detectors based on morphological erosion and hit-or-miss operators is generated randomly. From this population an optimized subset of features is selected using a novel application of genetic algorithms. The selected features are then used to initialize a generalized Hamming neural network that performs image classification. This network provides the means for self-organizing the set of training patterns into additional subclasses this in turn dynamically alters the number of detectors and the size of the neural network. The design process uses system errors to gradually refine the set of feature vectors used in the classification subsystem. We describe an experiment in which the hybrid learning paradigm successfully generates a machine that distinguishes ten classes of handprinted numerical characters.© (1990) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Optimal probabilistic relaxation labeling.

Abstract: This paper investigates the theoretical limits of probabilistic relaxation labeling (PRL), applied to per-pixel contextual image classification. The performance of a scheme which is defined to be optimal (within a class of PRL schemes) is studied, and found to fall short of that theoretically obtainable by directly considering all the original a posteriori probabilities (PPs) in the image. Lt is also found that an optimal scheme must use different updating functions at each iteration, and that these functions will depend on the distributions of the original per-pixel data.

A graphical approach for the evaluation of land-cover classification procedures

Abstract: A method for evaluating the effectiveness of different feature combinations and training strategies is described. Preliminary tests have been made using two groups of feature combinations derived from SPOT High Resolution Visible (HRV) data and two sets of training samples. The method is objective, and needs no ground confirmation or interaction from the image analyst. It is recommended as a surrogate for detailed accuracy assessment when attempting to find an optimum set of training pixels or feature combinations for image classification.

Textural image classification using variograms

Abstract: A method of image classification based on image texture is presented. Texture analysis is performed using the variogram function. The classifier used is a supervised parallelepiped type classifier using a minimum distance to mean check for overlapping classifications.© (1990) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Image classification by using associative noise-like coding memories

Abstract: The main goal of this research is to exploit the noise insensitivity and the pattern-completion capability shown by associative memories in order to improve pattern classification in the presence of noisy input data. A stochastic analysis of the adopted associative model is performed: the estimated noise variance is identified as a discriminating parameter for classification, and the optimal threshold value, which minimizes error probability in the filtering and presentation of recalls, is determined. Such theoretical results make it possible to outline the basic structure of an associative recognition system, in which a set of modules perform error-variance measurements and yield a candidate for recognition. A closed-loop version of such a system is also provided so that iterative associative reading cycles may improve the quality of recalls. The system can also be supported by techniques that make its behavior shift and scale independent; special effort is devoted to achieving the maximum modularity of the overall structure. Binary-image classification is performed as a testing task, and experimental results that prove the accuracy of theoretical predictions are reported. >

Selection of optimal textural features for maximum likelihood image classification

Abstract: The classification or segmentation of images into land cover or object class is one of the fundamental remote sensing/image processing tasks. The classification techniques have reached a high level of maturity in the forms of statistical pattern recognition techniques applied to multispectral images. According to this approach each pixel has a spectral vector associated with it and pixels are segmented into the class they most closely resemble spectrally. While these techniques are reasonably successful they do not take advantage of the brightness patterns within a class or at object boundaries nor are they readily applicable to monochrome images or highly correlated multispectral images (e.g., true color images). To overcome these limitations several investigators have suggested the use of image derived features as additional factors for use in image classification. Robert (1989) has identified over forty textural features that have been suggested by various authors as useful for scene classification. Regrettably it is very compute intensive to generate all these features for every pixel in an image and to then use them in a classifier. Schott et al. (1988) developed a technique for selecting a small subset of spectral bands from a large set based on criteria intended to optimize maximummore » likelihood classification. Salvaggio et al. (1990) have implemented code to generate 46 image derived textural features and applied a two-step feature reduction and optimization process based largely on the band selection process of Schott et al. (1988). The current effort drew on this proof-of-concept work mentioned above and improved on several limitations noted in the earlier work. Specifically, the initial feature set reduction algorithm of Robert (1989) is shown to be deficient and an improved method implemented. 18 refs., 62 figs., 13 tabs.« less

Image texture analysis : classification and segmentation.

Abstract: Texture is an ubiquitous surface characteristic of real-world scenes, and is a prominent clue in the human perception of the surrounding world. Texture analysis is essential to the analysis of many types of images. Classification of a given texture sample into one of a finite number of prototype textures (texture classification), and partitioning of a given image into disjoint regions of homogeneous textures (texture segmentation) are the two fundamental problems in texture analysis. This thesis is devoted to the development of effective and efficient algorithms towards the resolution of these two problems, and thus gaining a better understanding of texture. Existing approaches to texture analysis, being mostly ad hoc in nature, are mainly concerned with texture classification. For texture segmentation, very limited success has been reported in the open literature as it is a much more challenging problem. The first class of methods developed in this thesis is based on local vector mapping, a new concept introduced in the thesis. This concept enables one to offer a consistent and convenient explanation, a unification and generalization of some existing approaches. Functionally, a local vector mapping operation is composed of local neighborhood specification, local vector formation and vector mapping or transformation. Successful algorithms for both texture classification and segmentation are developed based on this concept. The second class of methods draws upon the recent findings in the physiology and psychophysics research communities. As the human visual system (HVS) is no doubt the most effective and efficient system to perform texture analysis, computational algorithms should mimic the HVS. One of the most popular models for the processing of pictorial information in the HVS is the so-called spatial channel model. This model hypothesizes that the input pictorial information to the visual cortex is processed via a set of parallel, quasi-independent channels each of which is tuned to a specific band of spatial frequency and orientation. By adopting this model, very successful algorithms for texture classification and segmentation are developed in this thesis. In addition to the two major classes of approaches mentioned above, the thesis also contains chapters on the description of new algorithms for feature extraction in texture classification based on local surface fitting and primitive analysis. These algorithms can be considered as supplementary to the first class of algorithms.

Feature Selection and Decision Space Mapping for Sensor Fusion

Abstract: An information fusion approach is presented for mapping a multiple dimensional feature space into a lower dimensional decision space with simplified decision boundaries. A new statistic, called the tie statistic, is used to perform the mapping by measuring differences in probability density functions of features. These features are then evaluated based on the separation of the decision classes using a parametric beta representation for the tie statistic. The feature evaluation and fusion methods are applied to perform texture recognition.

Analysis of x-ray hand images for bone age assessment

Abstract: In this paper we describe a model-based system for the assessment of skeletal maturity on hand radiographs by the TW2 method. The problem consists in classiflying a set of bones appearing in an image in one of several stages described in an atlas. A first approach consisting in pre-processing segmentation and classification independent phases is also presented. However it is only well suited for well contrasted low noise images without superimposed bones were the edge detection by zero crossing of second directional derivatives is able to extract all bone contours maybe with little gaps and few false edges on the background. Hence the use of all available knowledge about the problem domain is needed to build a rather general system. We have designed a rule-based system for narrow down the rank of possible stages for each bone and guide the analysis process. It calls procedures written in conventional languages for matching stage models against the image and getting features needed in the classification process.© (1990) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Image sequence coding using data classification

Abstract: In this paper a specific data classification approach is presented which can be used in coders for transmitting video sequences. The image data is locally classified into four categories prior to the coding process which is then divided in four modes in order to exploit the masking effect of motion and to reach high performance. The four categories are: moving area containing an edge (data class Cf) moving area containing no single edge (class C/j) still area containing an edge (class C) and still area containing no single edge (class CE). This classification approach is based on the fact that the human visual system is especially sensitive to edge information and this in function of the motion behavior. Generally it is known that the edge information is critical to the human eyes. In the moving areas however the human eyes are less critical of the spatial details in comparison to that in the still areas because of the masking effect of motion. The data in the moving areas can be treated relatively coarsely but when edges are observed the quality criterion should still be raised appropriately. Using this four-mode coding method the subjective reproduction quality can be improved even in the case of low bit-rate applications. This classification approach is applied in the conventional motion compensating hybrid (DCT/DPCM) coder although it can also be with other coders.© (1990) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Knowlege- Based Image Classification.

Abstract: In a knowledge based classification, ancillary data and knowledge are combined with spectral information. A method of knowledge based classification based on temporal relationships between classes is introduced. Knowledge about crop rotations is represented by means of state transition matrices. Spectral image information, information stored in a geographic information system and knowledge as represented in a matrix are combined in a Bayesian maximum likelihood classification. This method is elaborated for atest area in The Netherlands. Depending on the spectral separation of the classes and the level of detail of the transition matrices, the overall accuracy of the classification increased by 4 to 20% with respect to the result based on only spectral information.

A Biologically-Inspired Neural Network Architecture for Image Processing

Abstract: : This thesis project included a literature survey of biological and artificial neural network research followed by development and testing of high- order and image recognition hierarchical neural network algorithms Following training, performance testing of second-order and third-order networks yielded maximum accuracies comparable to those achieved by multilayer perceptron classifiers operating on test data sets Several versions of an image classification algorithm were tested for learning performance using pixel data from forward-looking infrared (FLIR) images of tanks, trucks, target boards, and clutter Employing the biologically-motivated Lambertization and contrast normalization of pixel windows, correlations with multiple Gabor function wavelets, and a 'phase synchronizing' local averaging routine, the image classification network extracted data features Different network versions fed the extracted features to varying output classification schemes To improve separation of problem classes, recommendations were made for varying the parameters of the Gabor function wavelets and modifying the phase synchronization scheme to extract more suitable features from image pixel data

Image pattern algorithms using neural networks

Abstract: The ability to classify texture regions in images is considered to be an important aspect of scene analysis. The information gained from such classification can be used by a computer vision system to assist in image segmentation as well as object identification. In this paper, the use of a neural network model in performing classification of images containing regular textures is investigated. The texture features used in the classification process are Hough transform-based descriptors. The performance and capabilities of the neural network approach are then compared to classical technique utilizing a linear associative memory.