Showing papers on "Contextual image classification published in 1995"

Texture segmentation using fractal dimension

Abstract: This paper deals with the problem of recognizing and segmenting textures in images. For this purpose the authors employ a technique based on the fractal dimension (FD) and the multi-fractal concept. Six FD features are based on the original image, the above average/high gray level image, the below average/low gray level image, the horizontally smoothed image, the vertically smoothed image, and the multi-fractal dimension of order two. A modified box-counting approach is proposed to estimate the FD, in combination with feature smoothing in order to reduce spurious regions. To segment a scene into the desired number of classes, an unsupervised K-means like clustering approach is used. Mosaics of various natural textures from the Brodatz album as well as microphotographs of thin sections of natural rocks are considered, and the segmentation results to show the efficiency of the technique. Supervised techniques such as minimum-distance and k-nearest neighbor classification are also considered. The results are compared with other techniques. >

An experimental comparison of autoregressive and Fourier-based descriptors in 2D shape classification

Abstract: An experimental comparison of shape classification methods based on autoregressive modeling and Fourier descriptors of closed contours is carried out. The performance is evaluated using two independent sets of data: images of letters and airplanes. Silhouette contours are extracted from non-occluded 2D objects rotated, scaled, and translated in 3D space. Several versions of both types of methods are implemented and tested systematically. The comparison clearly shows better performance of Fourier-based methods, especially for images containing noise. >

A detailed comparison of backpropagation neural network and maximum-likelihood classifiers for urban land use classification

Abstract: A detailed comparison of the backpropagation neural network and maximum-likelihood classifiers for urban land use classification is presented. Landsat Thematic Mapper images of Tucson, Arizona, and Oakland, California, were used for this comparison. For the Tucson image, the percentage of matching pixels in the two classification maps was only 64.5%, while for the Oakland image it was 83.3%. Although the test site accuracies of the two Tucson maps were similar, the map produced by the neural network was visually more accurate; this difference is explained by examining class regions and density plots in the decision space and the continuous likelihood values produced by both classifiers. For the Oakland scene, the two maps were visually and numerically similar, although the neural network was superior in suppression of mixed pixel classification errors. From this analysis, the authors conclude that the neural network is more robust to training site heterogeneity and the use of class labels for land use that are mixtures of land cover spectral signatures. The differences between the two algorithms may be viewed, in part, as the differences between nonparametric (neural network) and parametric (maximum-likelihood) classifiers. Computationally, the backpropagation neural network is at a serious disadvantage to maximum-likelihood, taking nearly an order of magnitude more computing time when implemented on a serial workstation. >

Local discriminant bases and their applications

Abstract: We describe an extension to the “best-basis” method to select an orthonormal basis suitable for signal/image classification problems from a large collection of orthonormal bases consisting of wavelet packets or local trigonometric bases The original best-basis algorithm selects a basis minimizing entropy from such a “library of orthonormal bases” whereas the proposed algorithm selects a basis maximizing a certain discriminant measure (eg, relative entropy) among classes Once such a basis is selected, a small number of most significant coordinates (features) are fed into a traditional classifier such as Linear Discriminant Analysis (LDA) or Classification and Regression Tree (CARTTM) The performance of these statistical methods is enhanced since the proposed methods reduce the dimensionality of the problem at hand without losing important information for that problem Here, the basis functions which are well-localized in the time-frequency plane are used as feature extractors We applied our method to two signal classification problems and an image texture classification problem These experiments show the superiority of our method over the direct application of these classifiers on the input signals As a further application, we also describe a method to extract signal component from data consisting of signal and textured background

A comparison of wavelet transform features for texture image annotation

Abstract: A comparison of different wavelet transform based texture features for content based search and retrieval is made. These include the conventional orthogonal and bi-orthogonal wavelet transforms, tree-structured decompositions, and the Gabor wavelet transforms. Issues discussed include image processing complexity, texture classification and discrimination, and suitability for developing indexing techniques.

Back-propagation network and its configuration for blood vessel detection in angiograms

Abstract: A neural-network classifier for detecting vascular structures in angiograms was developed. The classifier consisted of a multilayer feedforward network window in which the center pixel was classified using gray-scale information within the window. The network was trained by using the backpropagation algorithm with the momentum term. Based on this image segmentation problem, the effect of changing network configuration on the classification performance was also characterized. Factors including topology, rate parameters, training sample set, and initial weights were systematically analyzed. The training set consisted of 75 selected points from a 256/spl times/256 digitized cineangiogram. While different network topologies showed no significant effect on performance, both the learning process and the classification performance were sensitive to the rate parameters. In a comparative study, the network demonstrated its superiority in classification performance. It was also shown that the trained neural-network classifier was equivalent to a generalized matched filter with a nonlinear decision tree. >

Invariant matching and identification of curves using B-splines curve representation

Abstract: There have been many techniques for curve shape representation and analysis, ranging from Fourier descriptors, to moments, to implicit polynomials, to differential geometry features, to time series models, to B-splines, etc. The B-splines stand as one of the most efficient curve (surface) representations and possess very attractive properties such as spatial uniqueness, boundedness and continuity, local shape controllability, and invariance to affine transformations. These properties made them very attractive for curve representation, and consequently, they have been extensively used in computer-aided design and computer graphics. Very little work, however, has been devoted to them for recognition purposes. One possible reason might be due to the fact that the B-spline curve is not uniquely described by a single set of parameters (control points), which made the curve matching (recognition) process difficult when comparing the respective parameters of the curves to be matched. This paper is an attempt to find matching solutions despite this limitation, and as such, it deals the problem of using B-splines for shape recognition and identification from curves, with an emphasis on the following applications: affine invariant matching and classification of 2-D curves with applications in identification of aircraft types based on image silhouettes and writer-identification based on handwritten text

Decision-based neural networks with signal/image classification applications

Abstract: Supervised learning networks based on a decision-based formulation are explored. More specifically, a decision-based neural network (DBNN) is proposed, which combines the perceptron-like learning rule and hierarchical nonlinear network structure. The decision-based mutual training can be applied to both static and temporal pattern recognition problems. For static pattern recognition, two hierarchical structures are proposed: hidden-node and subcluster structures. The relationships between DBNN's and other models (linear perceptron, piecewise-linear perceptron, LVQ, and PNN) are discussed. As to temporal DBNN's, model-based discriminant functions may be chosen to compensate possible temporal variations, such as waveform warping and alignments. Typical examples include DTW distance, prediction error, or likelihood functions. For classification applications, DBNN's are very effective in computation time and performance. This is confirmed by simulations conducted for several applications, including texture classification, OCR, and ECG analysis. >

Combining image compression and classification using vector quantization

Abstract: We describe a method of combining classification and compression into a single vector quantizer by incorporating a Bayes risk term into the distortion measure used in the quantizer design algorithm. Once trained, the quantizer can operate to minimize the Bayes risk weighted distortion measure if there is a model providing the required posterior probabilities, or it can operate in a suboptimal fashion by minimizing the squared error only. Comparisons are made with other vector quantizer based classifiers, including the independent design of quantization and minimum Bayes risk classification and Kohonen's LVQ. A variety of examples demonstrate that the proposed method can provide classification ability close to or superior to learning VQ while simultaneously providing superior compression performance. >

Layout recognition of multi-kinds of table-form documents

Abstract: Many approaches have reported that knowledge-based layout recognition methods are very successful in classifying the meaningful data from document images automatically. However, these approaches are applicable to only the same kind of documents because they are based on the paradigm that specifies the structure definition information in advance so as to be able to analyze a particular class of documents intelligently. In this paper, the authors propose a method to recognize the layout structures of multi-kinds of table-form document images. For this purpose, the authors introduce a classification tree to manage the relationships among different classes of layout structures. The authors' recognition system has two modes: layout knowledge acquisition and layout structure recognition. In the layout knowledge acquisition mode, table-form document images are distinguished according to this. Classification tree and then the structure description trees which specify the logical structures of table-form documents are generated automatically. While, in the layout structure recognition mode, individual item fields in the table-form document images are extracted and classified successfully by searching the classification tree and interpreting the structure description tree. >

Classification of multisensor remote-sensing images by structured neural networks

Abstract: Proposes the application of structured neural networks to classification of multisensor remote-sensing images. The purpose of the approach is to allow the interpretation of the "network behavior", as it can be utilized by photointerpreters for the validation of the neural classifier. In addition, this approach gives a criterion for defining the network architecture, so avoiding the classical trial-and-error process. First of all, the architecture of structured multilayer feedforward networks is tailored to a multisensor classification problem. Then, such networks are trained to solve the problem by the error backpropagation algorithm. Finally, they are transformed into equivalent networks to obtain a simplified representation. The resulting equivalent networks may be interpreted as a hierarchical arrangement of "committees" that accomplish the classification task by checking on a set of explicit constraints on input data. Experimental results on a multisensor (optical and SAR) data set are described in terms of both classification accuracy and network interpretation. Comparisons with fully connected neural networks and with the k-nearest neighbor classifier are also made. >

A refined gamma MAP SAR speckle filter with improved geometrical adaptivity

Abstract: A modified version of the refined gamma maximum-a-posteriori (RGMAP) speckle filter, which is found in the literature, is presented. The traditional RGMAP speckle filter first defects contours belonging to step edges and thin linear structures, then applies the RGMAP filter to local statistics extracted from rectangular masks that do not cross image contours. The proposed modified RGMAP (MRGMAP) filter first exploits local operators belonging to the odd-symmetric filter category employed by RGMAP to detect image segments, then it computes local statistics over areas that are not necessarily rectangular, but are subsets of the image segments having any possible shape. Therefore, MRGMAP enhances the RGMAP ability in exploiting shape adaptive windowing near image contours, where speckle is not fully developed. The MRGMAP computation time is estimated to be of the same magnitude of that of the original RGMAP, the latter depending on the number of filter categories being employed. The qualitative and quantitative results of the MRGMAP filter applied to real SAR images are satisfactory as the filter seems to be effective in speckle removal whereas it retains edge sharpness and subtle details. However, tests on simulated SAR images must still be performed in order to provide definitive evidence supporting MRGMAP effectiveness. Since MRGMAP typically removes image structures featuring a constant reflectivity gradient, this filter is not particularly suitable for image enhancement in human photo-interpretation. MRGMAP can be rather employed as a preprocessing module in a computer-based SAR image classification procedure based on segment mean value analysis. >

Performance evaluation of texture measures for ground cover identification in satellite images by means of a neural network classifier

Abstract: The performance of several feature extraction methods for classifying ground covers in satellite images is compared. Ground covers are viewed as texture of the image. Texture measures considered are: cooccurrence matrices, gray-level differences, texture-tone analysis, features derived from the Fourier spectrum, and Gabor filters. Some Fourier features and some Gabor filters were found to be good choices, in particular when a single frequency band was used for classification. A Thematic Mapper (TM) satellite image showing a variety of vegetations in central Colorado was used for the comparison. A related goal was to investigate the feasibility of extracting the main ground covers from an image. These ground covers may then form an index into a database. This would allow the retrieval of a set of images which are similar in contents. The results obtained in the indexing experiments are encouraging. >

Feature selection in the pattern classification problem of digital chest radiograph segmentation

Abstract: In pattern classification problems, the choice of variables to include in the feature vector is a difficult one. The authors have investigated the use of stepwise discriminant analysis as a feature selection step in the problem of segmenting digital chest radiographs. In this problem, locally calculated features are used to classify pixels into one of several anatomic classes. The feature selection step was used to choose a subset of features which gave performance equivalent to the entire set of candidate features, while utilizing less computational resources. The impact of using the reduced/selected feature set on classifier performance is evaluated for two classifiers: a linear discriminator and a neural network. The results from the reduced/selected feature set were compared to that of the full feature set as well as a randomly selected reduced feature set. The results of the different feature sets were also compared after applying an additional postprocessing step which used a rule-based spatial information heuristic to improve the classification results. This work shows that, in the authors' pattern classification problem, using a feature selection step reduced the number of features used, reduced the processing time requirements, and gave results comparable to the full set of features. >

Experiments in curvature-based segmentation of range data

Abstract: This paper focuses on the experimental evaluation of a range image segmentation system which partitions range data into homogeneous surface patches using estimates of the sign of the mean and Gaussian curvatures. The authors report the results of an extensive testing program aimed at investigating the behavior of important experimental parameters such as the probability of correct classification and the accuracy of curvature estimates, measured over variations of significant segmentation variables. Evaluation methods in computer vision are often unstructured and subjective: this paper contributes a useful example of extensive experimental assessment of surface-based range segmentation. >

Accelerating fractal image compression by multi-dimensional nearest neighbor search

Abstract: In fractal image compression the encoding step is computationally expensive. A large number of sequential searches through a list of domains (portions of the image) are carried out while trying to find the best match for another image portion. Our theory developed here shows that this basic procedure of fractal image compression is equivalent to multi-dimensional nearest neighbor search. This result is useful for accelerating the encoding procedure in fractal image compression. The traditional sequential search takes linear time whereas the nearest neighbor search can be organized to require only logarithmic time. The fast search has been integrated into an existing state-of-the-art classification method thereby accelerating the searches carried out in the individual domain classes. In this case we record acceleration factors from 1.3 up to 11.5 depending on image and domain pool size with negligible or minor degradation in both image quality and compression ratio. Furthermore, as compared to plain classification our method is demonstrated to be able to search through larger portions of the domain pool without increased the computation time.

Classification and feature extraction of AVIRIS data

Abstract: The processing of Airborne Visible-Infrared Imaging Spectrometer (AVIRIS) data is discussed both in terms of feature extraction and classification. The recently proposed decision boundary feature extraction method is reviewed and then applied in experiments. Results of classifications for AVIRIS data from Iceland 1991 are given with emphasis on geological applications. The classifiers used include neural network methods and statistical approaches. The decision boundary feature extraction method shows excellent performance for these data. >

Image coding by block prediction of multiresolution subimages

Abstract: The redundancy of the multiresolution representation has been clearly demonstrated in the case of fractal images, but it has not been fully recognized and exploited for general images. Fractal block coders have exploited the self-similarity among blocks in images. We devise an image coder in which the causal similarity among blocks of different subbands in a multiresolution decomposition of the image is exploited. In a pyramid subband decomposition, the image is decomposed into a set of subbands that are localized in scale, orientation, and space. The proposed coding scheme consists of predicting blocks in one subimage from blocks in lower resolution subbands with the same orientation. Although our prediction maps are of the same kind of those used in fractal block coders, which are based on an iterative mapping scheme, our coding technique does not impose any contractivity constraint on the block maps. This makes the decoding procedure very simple and allows a direct evaluation of the mean squared error (MSE) between the original and the reconstructed image at coding time. More importantly, we show that the subband pyramid acts as an automatic block classifier, thus making the block search simpler and the block matching more effective. These advantages are confirmed by the experimental results, which show that the performance of our scheme is superior for both visual quality and MSE to that obtainable with standard fractal block coders and also to that of other popular image coders such as JPEG. >

Terrain objects, their dynamics and their monitoring by the integration of GIS and remote sensing

Abstract: Geometrical and thematic data about terrain objects stored in a geographical information system (GIS) can be kept up-to-date by using remote sensing (RS) data. Geometrical and thematic data can be extracted from the RS data by segmentation and classification techniques respectively. The possibilities and reliability of the information extraction from RS data can be improved by the use of ancillary data and knowledge about the terrain objects. Object classification and aggregation hierarchies can be used to describe relationships between terrain objects; the categorization of the different types of terrain object dynamics that is presented will be based partly on these hierarchical relationships. Object models will be applied in a case study in which both the field geometry (field boundaries) and the crop type of agricultural fields are updated from a Landsat TM image. For that purpose, a three-stage strategy has been developed. In the first stage, the results of an edge detection procedure are integrated with fixed geometrical data already contained in the GIS by using knowledge about the aggregation structure and shape of the fields. In the next stage, the crop type of the fields is determined by means of object-based classification. Finally, conditional merging is performed to solve the problem of oversegmentation. The resulting field geometry was found to agree for 87% with the field geometry as determined by a photo-interpreter. >

Decision trees for automated identification of cosmic-ray hits in hubble space telescope images

Abstract: We have developed several algorithms for classifying objects in astronomical images. These algorithms have been used to label stars, galaxies, cosmic rays, plate defects, and other types of objects in sky surveys and other image databases. Our primary goal has been to develop techniques that classify with high accuracy, in order to ensure that celestial objects are not stored in the wrong catalogs. In addition, classification time must be fast due to the large number of classifications and to future needs for on-line classification systems. This paper reports on our results from using decision tree classifiers to identify cosmic ray hits in Hubble Space Telescope images. This method produces classifiers with over 95% accuracy using data from a single, upaired image. Our experiments indicate that this accuracy will get even higher if methods for eliminating background noise improve.

Neural network based segmentation of magnetic resonance images of the brain

Abstract: Presents a study investigating the potential of artificial neural networks (ANN's) for the classification and segmentation of magnetic resonance (MR) images of the human brain. In this study, the authors present the application of a learning vector quantization (LVQ) ANN for the multispectral supervised classification of MR images. The authors have modified the LVQ for better and more accurate classification. They have compared the results using LVQ ANN versus back-propagation ANN. This comparison shows that, unlike back-propagation ANN, the authors' method is insensitive to the gray-level variation of MR images between different slices. It shows that tissue segmentation using LVQ ANN also performs better and faster than that using back-propagation ANN.

Genetic algorithms as a tool for restructuring feature space representations

Abstract: This paper describes an approach being explored to improve the usefulness of machine learning techniques to classify complex, real world data. The approach involves the use of genetic algorithms as a "front end" to a traditional tree induction system (ID3) in order to find the best feature set to be used by the induction system. This approach has been implemented and tested on difficult texture classification problems. The results are encouraging and indicate significant advantages of the presented approach.

Image classification apparatus

Abstract: An image classification apparatus includes an image input device for inputting image data. A filing device stores the input image data and performs a read, retrieval, or edit operation on the data in accordance with a predetermined instruction. A normalizing unit corrects variations in various image input conditions or variations in image input devices to make the input conditions in execution agree with those in learning. A feature extracting unit extracts a feature amount effective for classification. A classification determination unit performs a classification determination operation on the basis of the feature amount obtained by the feature extracting unit. A display unit synthesizes a result of the classification determination operation with the input image, and displays the synthesized result. A learning control unit controls the feature extracting unit and the classification determination unit on the basis of the predetermined learning data. Accordingly, the classification apparatus classifies a portion of the input image which is difficult to extract by binarization or three-dimensional display alone, and displays the classified portion in an easily recognizable visual form.

A system for machine-written and hand-written character distinction

Abstract: In applications of character recognition where machine-printed and hand-written characters are involved, it is important to know if the character image, or the whole word, is machine- or hand-written. This is due to the accuracy difference between the algorithms and systems oriented to machine- or handwritten characters. Obviously, this type of knowledge leads to the increase of the overall system quality. In this work a classification system is presented which reads a raster image of a character and outputs two confidence values, one for "machine-written" and one for "hand-written" character classes, respectively. The proposed system features a preprocessing step, which transforms a general uncentered character image into a normalized form, then the feature extraction phase extracts relevant information from the image, and at the end, a standard classifier based on a feedforward neural network creates the final response. At the end, some results on a proprietary image database are reported.

Scenic classification methods for image and video databases

Abstract: The problem of scenic image classification is presented in the paper. On considering the specific nature of this problem, we propose a statistically data-based method, the Hidden Markov Model, to solve this problem. We segment an image and use the sequence of segments as the definition of the image; we then train a HMM on a test set of sequences/images to establish a classification. We present preliminary results on the use of a 1D HMM for classification of images as either indoor or outdoor.© (1995) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Sea mine detection and classification using side-looking sonar

Abstract: Coastal Systems Station has developed an approach for automatic mine detection and classification. The Detection Density ACF Approach was created by integrating the adaptive clutter filter (ACF) developed by Martin Marietta, the specification of target signature suggested by Loral Federal Systems, and the Attracted-Based Neural Network developed at NSWC Coastal Systems Station with a detection density target recognition criterion. The Detection Density ACF Approach consists of eight steps: image normalization, ACF, selecting the largest ACF output pixels, convolving the selected pixels with a minesize rectangular window, applying a Bayesian decision rule to detect minelike pixels, grouping the minelike pixels into objects, extracting object features, and classifying objects as either a mine or a nonmine with a neural network. When trained on features extracted from 30 sonar images and tested against another 30 images, this approach demonstrates very good performance: probability of detection and classification (pdpc) of 0.84 with a false alarm rate of 1.4 false calls per image. A performance analysis study shows that the detection density ACF approach performs very well and significantly reduces the false alarm rate.© (1995) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Page segmentation and classification utilising a bottom-up approach

Abstract: This paper presents the use of analysing the connected components extracted from the binary image of a document page. Such an analysis provides a lot of useful information, and will be used to perform skew correction, segmentation and classification of the document. We present a new algorithm for determining the skew angle of lines of text in an image of a document with the advantage that it only performs one iteration to determine the skew angle. Experiments on over 30 pages show that the method works well on a wide variety of layouts, including sparse textual regions, mixed fonts, multiple columns, and even for documents with a high graphical content.

Method for finding and classifying scanned information

Abstract: A system that scans an image into a computer system and then locates left and right edges of information within each row of the image. The system defines top and bottom rows of information as first and last rows of a number of consecutive rows of the image that have a number of pixels between the left and right edges. The system linearizes the edge data, places a left boundary at the leftmost point of the linearized left edge, and places a right boundary at the rightmost point of the linearized right edge. Each pixel within the boundaries is analyzed to classify the pixel as white, light gray, dark gray, black, colored, or as having some color; a percentage of pixels in each class is created; and the image classification is determined using these percentages. The system displays the image and image classification, with the boundary superimposed over the image.

Segmentation and recognition of symbols within handwritten mathematical expressions

Abstract: An efficient on-line recognition system for symbols within handwritten mathematical expressions is proposed. The system is based on the generation of a symbol hypotheses net and the classification of the elements within the net. The final classification is done by calculating the most probable path through the net under regard of the stroke group probabilities and the probabilities obtained by the symbol recognizer based on hidden Markov models.

Application of neural networks for sea ice classification in polarimetric SAR images

Abstract: Several automatic methods have been developed to classify sea ice types from fully polarimetric synthetic aperture radar (SAR) images, and these techniques are generally grouped into supervised and unsupervised approaches. In previous work, supervised methods have been shown to yield higher accuracy than unsupervised techniques, but suffer from the need for human interaction to determine classes and training regions. In contrast, unsupervised methods determine classes automatically, but generally show limited ability to accurately divide terrain into natural classes. In this paper, a new classification technique is applied to determine sea ice types in polarimetric and multifrequency SAR images, utilizing an unsupervised neural network to provide automatic classification, and employing an iterative algorithm to improve the performance. The learning vector quantization (LVQ) is first applied to the unsupervised classification of SAR images, and the results are compared with those of a conventional technique, the migrating means method. Results show that LVQ outperforms the migrating means method, but performance is still poor. An iterative algorithm is then applied where the SAR image is reclassified using the maximum likelihood (ML) classifier. It is shown that this algorithm converges, and significantly improves classification accuracy. The new algorithm successfully identifies first-year and multiyear sea ice regions in the images at three frequencies. The results show that L- and P-band images have similar characteristics, while the C-band image is substantially different. Classification based on single features is also carried out using LVQ and the iterative ML method. It is found that the fully polarimetric classification provides a higher accuracy than those based on a single feature. The significance of multilook classification is demonstrated by comparing the results obtained using four-look and single-look classifications. >