This paper deals with the problems of detection of low- observable small-targets form sequence of IR images against structural background and non-stationary clutter.

In this work, we propose a unified formulation for correlation techniques in a theoretical framework based on group theory. Indeed, correlation on groups can be defined in terms of transformation groups, feature spaces and invariant measures. As an important example, we describe the correlation functions for the affine group and some of its subgroups. Algorithms are illustrated on underwater images to detect and identify specific objects (such as trawl traces) independently of their pose on the sea bottom.

It is a key technology to search or locate targets in target identification and tracking systems. The searching algorithm must make sure that the interesting target be captured. At present, in most target tracking systems, a template is specified at first then the target is locate by template matching, or the target is located manually. With the development of the system, it is very necessary to realize the automatic capture, identification and tracking of the target.

The background noise of the images from passive sensors normally is non-Gaussian, it is strong relativity in column direction. This paper will present an IR target's detection method using difference filter based on space difference to deal with such image data. From the simulations, we can find that this method is effective for the correlative background.

For HRR profile identification application, it is of importance to use as few as possible of matched-filters under the restriction of given probability of error and rejection. In this paper, a new method of target recognition called range-polarization multi-dimensional correlation matching classification is presented, which combine the polarization information with high resolution range information to adapt arbitrary pose angle of target, the optimal expressions are designed according to the demand of error an reject probabilities. The prosed method was performed in three steps: First, the theory of range- polarization multi-dimensional correlation matching was introduced. Then, multi-dimensional correlation matching expression were trained. Finally, some experimental results were given. The results of experiments demonstrated that his method has advantages of small operation amount and less matched filters.

The formation of 2D ISAR images for radar target identification hold much promise for additional distinguish- ability between targets. Since an image contains important information is a wide range of scales, and this information is often independent from one scale to another, wavelet analysis provides a method of identifying the spatial frequency content of an image and the local regions within the image where those spatial frequencies exist. In this paper, a multiresolution analysis wavelet method based on 2D ISAR images was proposed for use in aircraft radar target identification under the wide band high range resolution radar background. The proposed method was performed in three steps; first, radar backscatter signals were processed in the form of 2D ISAR images, then, Mallat's wavelet algorithm was used in the decomposition of images, finally, a three layer perceptron neural net was used as classifier. The result of experiments demonstrated that the feasibility of using multiresolution analysis wavelet for target identification.

A statistical feature based classifier is presented for high range resolution (HRR) radar aircraft identification. The match score method is utilized to correct the range migration of range profiles. The primary statistics required to determine are the peak location probability function and the peak amplitude probability density function. A single peak's belief can be obtained basing on them. The Dempster- Shafter evidence theory is used in multipeak fusion to get last belief, basing on which decision of identification are made. The experimental results from real radar aircraft data demonstrate that a high probability of correct identification can be obtained using the method proposed in this paper and fusion of 2-look can indeed improve the probability of correct identification.

Through analyzing the different relative depth of a 3D surface between the artificial target and the natural scene, a novel method of target detection based on wavelet transformation and geometric characteristics of 3D surfaces' texture is prosed. The wavelet Holder constant, which denote the relative distance between camera and targets, are calculated in a series of different multi-resolution images, and the targets are detected by calculating the slope of beeline. The target, which cannot be detected using groovy fractal Hurst exponents, can be detected perfectly using this method. The results show this method can improve the capability of anti-disturbance, provide accurate estimation and is also suitable for identifying specific targets in a complex background.

Multi-sensor information fusion plays an important pole in object recognition and many other application fields. Fusion performance is tightly depended on the fusion level selected and the approach used. Feature level fusion is a potential and difficult fusion level though there might be mainly three fusion levels. Two schemes are developed for key issues of feature level fusion in this paper. In feature selecting, a normal method developed is to analyze the mutual relationship among the features that can be used, and to be applied to order features. In object recognition, a multi-level recognition scheme is developed, whose procedure can be controlled and updated by analyzing the decision result obtained in order to achieve a final reliable result. The new approach is applied to recognize work-piece objects with twelve classes in optical images and open-country objects with four classes based on infrared image sequence and MMW radar. Experimental results are satisfied.

An approach to detect and track moving objects with a stationary camera is presented in this paper. The mixture Gaussian model is used as an adaptive background updating method. Based on subtraction foreground is separated from background, and then foreground objects are segmented with a modified binary connected component analysis. Kalman filtering is used in object tracking. To deal with problems caused by occlusions between objects in tracking, six representative categories are introduced and analyzed. Experiments on several outdoors video streams resulted with convictive object detection and tracking performance demonstrate its strong adaptability to lighting changes, shadows and occlusions.

The data alignment techniques for Radar/IIR dual-sensor integration in some dual-module homing seeker will bee considered.

IN this paper, the advantage and disadvantage of several seal imprint recognition methods have been compared, and an approach for seal imprint recognition based on ring- projecting template matching is proposed. Firstly, the seal important in divided into some rings, and the local and global intrinsic features are extracted. The, the recognition is made, with the border and the words of seal imprint separated, by 1D template matching. The experimental results show that the proposed approach is of adaptability and reliability.

This paper presents a vision-based traffic scenes analysis system. We have developed a feature-based tracking approach of vehicle tracking. This important task is to group features that come from the same vehicle. The motion of features is detected by straight lines matching and motion vectors are reprojected to road to get reprojected velocities. According to the vehicle model, the vehicle height is presumed to estimate other features' height. Then the vehicle structure can be rebuilt by the relationship of feature's height and reprojected velocity. The rebuilt vehicle structure is verified by 3D model rules.

A ship detection algorithm used in a navigation lock monitored control system, which helps to improve safety and dependability, is presented. We mean, in this paper, the processed region of river channel is determined through image edge detection. A method of peak-cutting histogram is introduced to restrain illusive movement information resulting from the glisten of water in the navigation lock. A concept of difference compactness from computing the projected density function of difference image is put forward, which helps to detect moving ships more precisely. With the statistical property of histogram of several small regions of river channel: gray scale variance, the detection of stationary ships can be accomplished, and the results can be judged by confidence region. Computer simulation has proved that it contributes a lot to the detection of ships in a navigation lock.

The term crease is used to denote the total line features in high resolution palmprint images. In the bibliography, line features in low resolution palmprint images have been used in palmprint recognition. The result is limited due to the lack of lines in the poor images. In the high resolution situation, the creases are abundant while they also need much more deliberate treatment. A palmprint recognition system have been constructed in our previous research, which demonstrates the feasibility of palmprint recognition using only crease while it doesn't take into account many low quality images in reality. In this paper, we advance the work in image preprocessings, automatic threshold selection in the key part of system, and crease lines matching. Now the crease extraction and palmprint recognition results are exciting even in very low quality palmprint. Based on our system, an absolute threshold of similarity degree, which ensures all the image pairs with similarity degree above it are of the same palmprint, like that of matching point number in the fingerprint recognition, is figured out. All the results show that creases in high resolution palmprint images will be an outstanding candidate in the future palmprint recognition system with very high sample capacity.

Numerous methods have been applied in automatic target recognition (ATR) systems now, and a lot of factors can impact the system robustness and the recognition ratio. There is always a need for a system structure to adapt different recognition algorithms and provide rapid performance evaluation and comparisons of these algorithms. In this paper, a hierarchical modular structure for automatic target recognition system is brought forward. With the applying of the hierarchical target recognition method, a complicated multi-class recognition problem can be disintegrated into simpler recognition problems with fewer classes in different layers. Meanwhile, an ATR system with the modular structure consists of relatively stand-alone modules that implement different functions such as data acquisition, feature extraction and target recognition. Different algorithms can be implemented, providing optional modules for the system construction. Designers can easily choose between algorithms, adjust and optimize these modules respectively and provide the optimal design for the whole system.

A new correlation algorithm for tracking extended target in complex environment is proposed in this per. The moment invariants are chosen as the tracking feature since the size, shape and orientation of the extended target would change distinctly during movement. The improved dynamic threshold segmentation is introduced to remove edges of the background within the tracking window. Then we calculate the moment invariants based on edge region of the local target, through which the computation quality is reduced drastically. Based on these moment invariants, a new fast correlation algorithm is proposed, in which the multistage and subsample are adopted to improve the matching efficiency. Experimental results indicated the validity and stability of our algorithm for tracking the extended target in the conditions of complex environment, high noise, rotation and size variance.

The movement of an object can be detected by subtracting the successive two frames of the object's video sequence. On the image after subtraction, there is a larger difference in the object area while a smaller one in the background part. So the track of a moving object can be detected by calculating the position of its gravity center on the image. But if there are several moving objects or the background is moving too, the gravity center of the image is not the real one of any objects. In this paper, we divided the image into 16 by 16 equals 256 parts and calculate the gravity center of each one. Then we get rid of parts with small average weight according to a certain threshold. The parts belong to the same moving object should be grouped in space, and a new gravity center is obtained. It is quite easy to get rid of the parts belong to the background because their differences are small and their moving vector are same. On the condition of 768 by 576 resolution and 25f/s video frequency, we are able to detect the tracks of several moving objects against a moving background and display them on a PC screen in real time.

In this paper, a hybrid optoelectric processor based on morphological algorithm is described for Image Tracking. It is set up based on an incoherent optical correlator, which has some merits in compactness, immunity to coherent noise and flexibility for applications. A modified edge detection, a dilation operation and a two-dimensional extensive complementary encoding method are proposed to preprocess the images. An automatic postprocessing for improving the recognition capability is introduced. The scene image recognition and tracking experiments are implemented, and the experimental results are given.

In the model-based recognition methods, the result's confidence is decided by the feature distance between the segmented region and the target model, and can be defined as the posterior probability that can be computed from the object and background's prior probability and conditional probability with Bayesian formula. However, when recognizing the target, many physical constrains, or image measurements of object region and background region, can be applied on the validation of the recognition result, and should be introduced into the confidence analysis. In this paper, we proposed a new method to analyze the target recognition quality by combining the physical constrains or prior knowledge into confidence analysis within the frame of mathematical statistic theory and Dempster-Shafer's evidence theory. In this method, the usability of the information sources is appraised with Kolmogorov-Smirnov test method and the different computation models to compute the belief value to classifier's result corresponding to the different information source types were also proposed. The method was tested on the real sequences of images, and the result indicated that the proposed method for confidence analysis is feasible and effective.

We present a new algorithm that can detect and track a number of small targets or point targets with low contrast through an image sequence taken from a static camera. The issues that we have addressed to achieve this are twofold. Firstly, the detection of small targets or point targets based on order morphology filtering, and secondly, tracking of targets based on image flow analysis. The experiment results show that the method can effectively and reliably detect and track moving small targets or point targets with low SNR comprising high-pass filtering and Kalman track filtering.

We present a possible way to detect 3D out-of-plane targets. Several Su-27 airplane images with different 3D rotational views were used to synthesize a template function, which successfully detects the target against the non-target such as the F16 airplane. A theoretical development for the purpose of pattern recognition is proposed. The system has the desirable property of sharp peaks with low sidelobes in the output correlation plane when multiple targets appear in the input. The test results show that the correlation peak is quite distinguishable at the location of the target and indicate the success of the technique. When combining the advantages of optics and electronics, the system is suitable for hybrid optical/electrical signal processing.

There are tow respects to be studied on the characteristic of the laser echo of the underwater target in the paper. One is how to detect the echo signal of the target, and the other is how to describe it.

The classification of ship targets using the kernel Fisher discriminant analysis is investigated in this paper. The main idea of this method is to find a nonlinear direction by first mapping the data nonlinearly into some feature space and compute Fisher's linear discriminant in input space. Based on the kernel Fisher discriminant, we recognize three types of ships. The satisfactory experimental results are obtained. In addition, we compare this method with other state of the art classification techniques. The experiments show that the kernel Fisher discriminant is superior to the other algorithms.

According to the character for the point target and the weak and small target of the IR image, the method discerning and detecting the IR weak and small point target is discussed in this paper. The method of the background adaptive forecast is taken to catch the point target and the weak and small target. IT provides the accurate tracking of the IR image target with a good foundation. Meantime the theory of the adaptive control is utilized for analyzing the tracking and control system of IR image emphatically. According to the character and demand of the control system an adaptive fuzzy PID controller for the parameter is researched and designed to realize the accurate track of the point target and the weak and small target of the IR image. Through a large number of the imitation for the practical system the results prove that the adaptive controller is provided with the high controlling precision, high speed response and the satisfactory robustness.

The matching process between the input data of real hand's characteristic marks and postulated positions of the model is needed while using model-based methods in gesture recognition. This fitting process can be regarded as a procedure of optimization based on a certain hand alignment measure. In this paper, an improved algorithm for the model fitting process is proposed to substitute conventional searching algorithm. The algorithm is the combination of Genetic Algorithm (GA) and classical Line Searching Algorithm (LSA). The new algorithm guarantees the fitting solution to be the global optimal and the operation near to be real time implementable. The experimental results show that the proposed algorithm is capable of searching optimal internal model positions with high speed and precision.

The weighting exponent m is an important parameter in fuzzy c-means (FCM) algorithm. In this paper, three basic problems about m in FCM algorithm: clustering validity method based on optimal m (or whether does optimal m exist), how does m effect on the performance of fuzzy clustering, and which is the proper range of m in general applications, are studied with the knee of objective function Jm, and fuzzy decision-making methods. Numerical experimental results show that the optimal m* for specific data set does exist. Moreover, a group of numerical experimental results indicate that, within the range of m (epsilon) (1.5, 3.5), the optimal m* monotone increase linearly against the separability (rho) of data set. So in practical applications, one can choose the value of m within the range of [1.5, 3.5].

Hyperspectral remote sensing image classification generally adopts a direct spectral matching method. It is, however, inconvenient in the classification calculation because the complete reference spectra are needed. In this work we have developed a new chromaticity difference-based classification algorithm, which can be used to classify imaging spectrometer image data. In calculation, the algorithm itself is not directly relating to the number of spectral wavebands. It only needs three chromaticity coordinate parameters for both the image spectrum and the reference spectrum to complete the final classification calculation. In addition, the classification threshold for the algorithm can be easily set according to the color science theory, therefore, the classification results from the algorithm is reliable. Through a comparison with SAM algorithm, the performance of the new chromaticity difference-based classification algorithm was proved to be as good as SAM algorithm, but our algorithm was relatively simpler and flexible.

Image textures have been playing an important role in image recognition and interpretation. The selection of image texture features is the key part to classify image objects. An evolution approach for texture image feature selection is prosed in this paper. The approach uses evolution algorithms as the primary search component. Based on selected texture features, a fuzzy cluster method is used for texture image classification. Experimental result on color aerial images show the feasibility of the proposed method.

Combination of many different classifiers can improve classification accuracy. Sugeno and choquet integrals with respect to the fuzzy measure possess many desired properties, so in this paper they are used to combine multiple neural network classifiers. However, it is difficult to determine fuzzy measures in real problems. In this paper, we present two methods, one is that we assign the degree of importance of each network based on how good these networks classify each class of the training data, the other is by genetic algorithms (GAs), to obtain fuzzy measures, each taking into account the intuitive idea that each classifier always possesses different classification ability for each class. In the experiment, several databases in UCI repository are tested using these combination schemes and compared with C4.5. They are also applied to a multisensor fusion system for workpiece identification. Experimental results confirm the superiority of these presented methods.

We compared different edge detection algorithms and found that the 'Marr-Hildreth' edge detection have the best performance for macrocalcification shape preservation in our MCCs detection system. Edge detection is one of the most commonly used operations in image analysis. The edges form the outline of the macrocalcifications. An edge is the boundary between an object and the background, and indicates the boundary between overlapping objects. This means that if the edges in an image can be identified accurately, all of the macrocalcifications can be located and the areas, perimeters, and shapes can be measured. So edge detection is one of the effect methods to preserve the shape of microcalcifications. Based on the edge enhancement method, a new mixed feature multistage method has been developed for improving the false positive (FP) reduction performance. Eleven features were extracted from both spatial and morphology domains in order to describe the micro-calcification clusters (MCCs) from different perspectives. These features are grouped into three categories: gray-level description, shape description and clusters description. This method was combined with neural network used in our false positive reduction, that reduce the false positive from 3.1/image to 0.1/image in 50 full field digital mammograms, The 50 mammograms are with 24 normal images and 26 abnormal images, including 41 microcalcification clusters in our database.

Association rule clustering is one of the most important topics in data mining. This paper proposes a generalization of distance-based clustering algorithm of association rules on various types of attributes. Firstly, considering complex database with various data, we present numeralized processing to deal with rules on many kinds of attributes. Secondly, instead of these values of numerilzed attributes being computed straightly, we propose an approach to normalized these attributes of association rules. Finally, with applying the numeralized as well as normalization methods, we present the generalization of clustering algorithm based on the different definitions of distances and diameters of rules. This algorithm can be used to handle the rules with attributes of different types and different scales, which extend the method of clustering. Tow simple examples are also provided to demonstrate the better result of the clustering algorithm in the end of the paper.

In this paper, we present a novel approach to the classification of texture images in the JPEG domain using the hard and soft thresholding functions, avoiding the need of the various stages of the decompression. A texture block of 8X8 size in discrete cosine transform (DCT) form is assigned to the most similar and nearest cluster center, where the shortest distance is selected from the list of distances of the blocks of the texture from the different cluster centers, that have been already calculated using fuzzy learning vector quantization (FLVQ), in which means and variances of AC energy of DCT blocks are inputs.

A target classification algorithm in high resolution and single polarimetric SAR image is presented in this paper. First, a new RCS reconstruction filter base don the modified correlated neighborhood model is used for target and shadow detection. By nonlinear iterated process, the whole image can be classified to 'background', 'shadow' and 'targets'. Secondly, a series of morphological operators are used to terrain filtering and edge extraction, and then through modified Hough Transform technology and the measure of slimming lines, line structures of man-made clutters in 'target' class are linked and grouped. Finally, with spatial association mode, the targets we are interested in are classified.

Based on local features, a new adaptive classification approach for multispectral remote sensing data is presented. Typical classification techniques based on global features tend to degrade because all classes are projected along the same direction, e.g. the principal component direction for Principal Component Analysis (PCA) and minimum component direction for Minimum Component Analysis (MCA). The typical methods are under the assumption that class separability is uniform for all directions, which is not always true. The new method overcomes that disadvantage by selecting features, which give the maximum class separability, based on local information of the classes instead of global information. In the new method, a projection matrix for every class is first sought based on making its training examples well separated from the others. Every input vector is then linearly transformed into another space by every projection matrix. In the transformed spaces, it can be classified or labeled to different class by Maximum Likelihood Classification (MLC). In order to reduce computation cost, adaptive dimension reduction is also introduced. Good performance of the new method can be shown from the experimental results on the Kennedy Space Center (KSC) TM images.

In this paper, we attempt to argue that the uncertainty coming form fuzzy information is ubiquitous in an intelligent fault diagnosis system and that fuzzy pattern recognition is an appropriate tool for the diagnosis of faults in complex devices. In the first place, the characteristics of the faults in a complex equipment system are introduced along with the fuzzy pattern recognition method and principle in intelligent fault diagnosis systems. Then, on the base of the above discussion, the paper gives an applied approach to fault diagnosis that combines the valve value rule with the maximum membership degree rule. Lastly, the practicability and validity of the method is illustrated through a practical example.

A multiple neural network classifier fusion system design method using immune genetic algorithm (IGA) is proposed. The IGA is modeled after the mechanics of human immunity. By using vaccination and immune selection in the evolution procedures, the IGA outperforms the traditional genetic algorithms in restraining the degenerate phenomenon and increasing the converging speed. The fusion system consists of N neural network classifiers that work independently and in parallel to classify a given input pattern. The classifiers' outputs are aggregated by a fusion scheme to decide the collective classification results. The goal of the system design is to obtain a fusion system with both good generalization and efficiency in space and time. Two kinds of measures, the accuracy of classification and the size of the neural networks, are used by IGA to evaluate the fusion system. The vaccines are abstracted by a self-adaptive scheme during the evolutionary process. A numerical experiment on the 'alternate labels' problem is implemented and the comparisons of IGA with traditional genetic algorithm are presented.

Stellar spectra classification is an indispensable part of any workable automated recognition system of celestial bodies. Like other celestial spectra, stellar spectra are also extremely noisy and voluminous; consequently, any acceptable technique of classification must be both computationally efficient and robust to structural noise. In this paper, we propose a practical stellar spectral classification technique which is composed of the following three steps: In the first step, the Haar wavelet transform is used to extract spectral lines, then followed by a de-noising process by the hard thresholding in the wavelet field. As a result, in the subsequent steps, only those extracted spectral lines are used for classification due to the high reliability of spectral lines with respect to the continuum. In the second step, the Principal Component Analysis (PCA) is employed for optimal data compression. More specifically, we use 165 well-selected samples from 7 spectral classes of stellar spectra to construct the 'eigen-lines spectra' by PCA. Thirdly, unknown spectra are projected to the eigen-subspace defined by the above eigen-lines spectra, and then a fuzzy c-means algorithm is used for the final classification. The experiments show that our new technique is both robust and efficient.

Fractional Brownian motion (FBM) plays an important role in the study of fractal characteristics of images. However, because FBM is an isotropic model, it cannot reflect the directional information of fractals. There is also scale range limit of the fractal dimension due to limitations in the FBM model. In this paper, an extended FBM is proposed to involve directional fractal information. Furthermore, an interception-and-linearity (IAL) method is presented to determine the cutoff scales of the fractal dimension automatically instead of choosing them through human observation. Textures with directionality are analyzed properly through the extended FBM method compared with the original one. For the purpose of classification, the fractal dimensions concerning directional information are obtained through the proposed method of scale range determination. Results have shown correct and subtle classification of textures.

Rotational invariant texture classification is required for many real world applications. Rotation invariant texture features are derived from the even symmetric Gabor filtered images of texture. The feature used is ADD from mean. It can be shown that rotation of input image is equivalent to a translation of the channel output along the orientation axis. This property is exploited to convert rational variant features to rotational invariant features. Discrete Fourier Transform of the feature is taken in rogation dimension to make the feature ration invariant. The classification of 45 Brodatz textures rotated in 12 different directions is done using these features. The number of samples used for training and testing phase are 4320. The percentage correct classification is 85.25.

This paper describes a high-speed coarse classifier, which makes use of a variable candidate selection method. The classifier is applicable to large character set recognition, such as Chinese, Japanese character. In designing the classifier, three strategies are used: lookup table, dimension reduction, and variable number of candidate selection. The classifier points to two directions: speeding up candidate selection and reduce the candidate set as much as possible. Compared with the fixed number candidate selection method, the third strategy can reduce the average candidate length significantly. In addition, we proposed an adaptively threshold estimating algorithm using distance histogram. The performance of this coarse classifier was test on the 863 Testing System. Experimental results verified its affectivity.

In this paper, a method of sub-classes division based on statistical and discrimination based on structural feature information is proposed. Some necessary structural feature and/or some micro feature are employed for further categorization in addition to these measures. The orientation and its variation of an entire image, and branch stroke trace, its orientation and its variation, the length and its variation are also considered. The first rough classification is mainly based on macro structural feature. In the second stage on ly forty-one cluster prototypes are built, according to statistical distribution. A modified ISODATA clustering is designed for the third classifier. Using the structural difference with some necessary/sufficient feature information and new normalization means made fine discrimination more exact and rapid, avoiding too many primitives and too long sentences and too many grammars for final classification.

D-S evidence theory is a useful method in dealing with uncertainty problems, but its application is limited because of the shortcomings of its combination rule. This paper present an efficient combination rule, that is, the evidences' conflicting probability is distributed to every proposition according to its average supported degree. The new combination rule improves the reliability and rationality of combination results. Although evidences conflict one another highly, good combination results are also obtained.

Level set method is a numerical analysis tool for the computation of propagating interfaces. In order to accelerate and strengthen numerical implementations of level set method, this paper proposed a set of auxiliary algorithms, i.e., simplified scanning based distance function construction, curve inside and outside labeling algorithm, and scanning based extension of velocity. With the former two algorithms, the signed distance function used as level set function can be constructed faster and more accurate. And the last algorithm makes the curve evolution by level set method more robust. The numerical tests showed efficiencies of the proposed algorithms.

Recently, with the development of multimodal user interfaces, handwritten Chinese character recognition, as an important part of handwriting input modality, has been paid particular attention. Researchers have done a large number of works in this field. Especially in the 1990s, numerous theories and methods have been developed. In this paper, the development of handwritten Chinese character recognition reported since 1990 is reviewed, and its applications of it in multimodal are also discussed.

A modified method for performing nonlinear form of Fuzzy C-Means (FCM) clustering algorithm (K-FCM) is proposed. By the use of kernel mapping, the non-linear clustering problem can be efficiently transformed into a linear problem in high-dimensional, even infinite, feature space. At the same time, we need not to know the explicit form of the non-linear mapping. That means that the computational complexity will not raised largely. The experimental result reveals the efficient and effective of the method proposed in this paper.

Color image segmentation is very important to machine vision, image understanding, and content-based image retrieval, etc, but there are few automatic and effective algorithms that can work fast and well on natural scene image. In this paper, we propose a new automatic color image segmentation algorithm using perceptually color clustering in Munsell(HVC) color space. Above all, we introduce the conversion formulae from (R, G, B) to (H, V, C) and the NBS color distance. Based on this, first, colors in the image are quantized to 256 colors or fewer without significantly degrading the color quality; then clustering the similar colors based on NBS color distance, finally, according to some rule, merging small color regions to its neighbor region.

Fingerprint recognition is an important subject in biometrics to identify or verify persons by physiological characteristics, and has found wide applications in different domains. In the present paper, we present a finger recognition system that combines singular points and structures. The principal steps of processing in our system are: preprocessing and ridge segmentation, singular point extraction and selection, graph representation, and finger recognition by graphs matching. Our fingerprint recognition system is implemented and tested for many fingerprint images and the experimental result are satisfactory. Different techniques are used in our system, such as fast calculation of orientation field, local fuzzy dynamical thresholding, algebraic analysis of connections and fingerprints representation and matching by graphs. Wed find that for fingerprint database that is not very large, the recognition rate is very high even without using a prior coarse category classification. This system works well for both one-to-few and one-to-many problems.

In this paper, the concept of 'interactive extraction' is discussed. A practical semiautomatic system should attend four factors: correctness, robustness, accuracy and interactivity. The paper mainly specifies the adjustment model of object-space based interactive extraction of manmade object from aerial image pair. Through the input points that donate the rough position of the object, it is matched with the object model by the least square template matching, so the corrected parameters describe the object and the coordinates of the object are acquired in object-space directly. A linear object can be described as a sample line in object-space. By matching the edge or road profile template with the initial input curve, it is derived that the adjustment model which expresses the correction of the curve parameters as the erroneous equations between the gray level of the pixel and the best-matched template. To evaluate the ground coordinates of the building corners, the adjustment model is defined as straight edge matching with object-space based geometric constraints. The object-space based geometric model is a flexible framework of extraction of manmade object from space image. The experimental results indicate that the method is ready for practical production.

Generally, our detection technique detects man-made targets of interest when those targets are a small percentage of a nature background. If we use one threshold for the integral field, some nature background would be divided to wrong kind of areas. In this paper, the intrinsic difference in fractal features between natural background and man-made objects is presented. By the discussion of its characteristic using wavelet composition, the fractal dimension and the error of fractal feature after multiscale decomposing is used to detect man-made targets in natural background. Our experimental results for real images show that the procedure in this paper is applicable.

In this paper we present a coherent image processor, which uses the loop-like nonlinear diffraction property of the Ce:KNSBN photorefractive crystal to achieve real-time edge- enhanced optical correlation and optical pattern recognition. The full-width-at-half-maximum of the auto- correlation peak is decreased four times by the edge- enhancement operation. The intensity ratio between auto- correlation peak and the cross-correlation peak is improved nearly two times. As the same time the lobe of the auto- correlation peak and the background 'clutter' noise are obviously suppressed. Therefore the discrimination of this system achieves great improvement.

In geographic space, it is well known that spatial behaviors of humans are directly driven by their spatial cognition, rather than by the physical or geometrical reality. The cognitive distance in spatial cognition is fundamental in intelligent pattern recognition. More precisely, the cognitive distance can be used to measure the similarities (or relevance) of cognized geographic objects. In the past work, the physical or Euclidean distances are used very often. In practice, many inconsistencies are found between the cognitive distance and the physical distance. Usually the physical distance is overestimated or underestimated in the process of human spatial behaviors and pattern recognition. These inconsistencies are termed distance distortions. The aim of this paper is to illustrate the conceptions of cognitive distance and distance distortion. And if the cognitive distance is argued to be two-dimensional, it exists in heterogeneous space and the property of quasi-metric is shown. If the cognitive distance is multi-dimensional, it exists in homogeneous space and the property of metric is shown. We argue that distance distortions arise from the transformation of homogeneous to heterogeneous space and from the transformation of the two-dimensional cognitive distance to the multi-dimensional cognitive distance. In some sense, the physical distance is an instance of cognitive distance.