With the advent of high-definition television, video phone, Internet and video on PCs, media content has to be displayed with different resolutions and high quality image interpolation techniques are increasingly demanded. Traditional image interpolation methods usually use a uniform interpolation filter on the entire image without any discrimination and they tend to produce some undesirable blurring effects in the interpolated image. Some content adaptive interpolation methods have been introduced to achieve a better performance on specific image structures. However, these content adaptive methods are limited to fitting image data into a linear model in each image structure. We propose extending the linear model to a flexible non-linear model, such as a multilayer feed-forward neural network. This results in a new interpolation algorithm using neural networks with coefficients based on the known pixel classification. Due to the fact that the number of classes using the pixel classification increases exponentially along with the filtering aperture size, we further introduce an efficient method to reduce the number of the classes. The results show that the proposed algorithm demonstrates more robust estimation in image interpolation and gives an additional improvement in the interpolated image quality. Furthermore, the work also shows that the use of pre-classification limits the complexity of the neural network while still achieving good results.

The exploration and categorization of large and unannotated image collections is a challenging task in the field of image retrieval as well as in the generation of appearance based object representations. In this context the Self-Organizing Map (SOM) has shown to be an efficient and scalable tool for the analysis of image collections based on low level features. Next to commonly employed visualization methods, clustering techniques have been recently considered for the aggregation of SOM nodes into groups in order to facilitate category specific data exploration. In this paper, spectral clustering based on graph theoretic concepts is employed for SOM based data categorization. The results are compared with those from the Neural Gas algorithm and hierarchical agglomerative clustering. Using SOMs trained on an eigenspace representation of the Columbia Object Image Library 20 (COIL20), the correspondence of the cluster data to a semantic reference grouping is calculated. Based on the Adjusted Rand Index it is shown that independent from the number of selected clusters, spectral clustering achieves a significantly higher correspondence to the reference grouping than any of the other methods.

In this paper, we propose an efficient similarity search system PathSOM that combines Self-Organizing Map (SOM) and Pathfinder Networks (PFNET). In the front end of the system, SOM is applied to cluster the original data vectors and construct a visual map of the data. The Pathfinder network then organizes the SOM map units in the form of a graph to yield a framework for an improved search to find the best matching map unit. The ability of PathSOM approach for efficient searches is demonstrated through well-known data sets.

There exist various objects, such as pictures, music, texts, etc., around our environment. We have a view for these objects by looking, reading or listening. Our view is concerned with our behaviors deeply, and is very important to understand our behaviors. We have a view for an object, and decide the next action (data selection, etc.) with our view. Such a series of actions constructs a sequence. Therefore, we propose a method which acquires a view as a vector from several words for a view, and apply the vector to sequence generation. We focus on sequences of the data of which a user selects from a multimedia database containing pictures, music, movie, etc... These data cannot be stereotyped because user's view for them changes by each user. Therefore, we represent the structure of the multimedia database as the vector representing user's view and the stereotyped vector, and acquire sequences containing the structure as elements. Such a vector can be classified by SOM (Self-Organizing Map). Hidden Markov Model (HMM) is a method to generate sequences. Therefore, we use HMM of which a state corresponds to the representative vector of user's view, and acquire sequences containing the change of user's view. We call it Vector-state Markov Model (VMM). We introduce the rough set theory as a rule-base technique, which plays a role of classifying the sets of data such as the sets of “Tour”.

Erythropoietin (Epo) is a hormone which can be misused as a doping substance. Its detection involves analysis of images containing specific objects (bands), whose position and intensity are critical for doping positivity. Within a research project of the World Anti-Doping Agency (WADA) we are implementing the GASepo software that should serve for Epo testing in doping control laboratories world-wide. For identification of the bands we have developed a segmentation procedure based on a sequence of filters and edge detectors. Whereas all true bands are properly segmented, the procedure generates a relatively high number of false positives (artefacts). To separate these artefacts we suggested a post-segmentation supervised classification using real-valued geometrical measures of objects. The method is based on the ID3 (Ross Quinlan's) rule generation method, where fuzzy representation is used for linking the linguistic terms to quantitative data. The fuzzy modification of the ID3 method provides a framework that generates fuzzy decision trees, as well as fuzzy sets for input data. Using the MLT^TM software (Machine Learning Framework) we have generated a set of fuzzy rules explicitly describing bands and artefacts. The method eliminated most of the artefacts. The contribution includes a comparison of the obtained misclassification errors to the errors produced by some other statistical classification methods.

In this paper, we present a new system to segment and label the contents of scanned documents as either text or image, using a modified fuzzy c-means (FCM) algorithm. Each pixel is assigned a feature pattern extracted from the gray level distribution and computed at different scales. The invariant feature pattern is then assigned to a specific region using fuzzy logic. Our algorithm is formulated by modifying the objective function of the standard FCM algorithm to allow the labeling of a pixel to be influenced by the labels in its immediate neighborhood. The neighborhood effect acts as a regularizer and biases the solution towards piecewise-homogeneous labelings. Such a regularization is useful in segmenting scans corrupted by scanner noise.

Offline handwritten chinese character recognition (HCCR) is still a difficult problem because of its large stroke changes, writing anomaly, and the difficulty for obtaining its stroke ranking information. Generally, offline HCCR can be divided into two procedures: feature extraction for capturing handwritten Chinese character information and feature classifying for character recognition. In this paper, we proposed a new chinese character recognition algorithm. In feature extraction part, we adopted elastic mesh dividing method for extracting the block features and its relative fuzzy features that utilized the relativities between different strokes and distribution probability of a stroke in its neighbor sub-blocks. In recognition part, we constructed a classifier based on a supervisory competitive learning algorithm to train competitive learning neural network with the extracted features set. Experimental results show that the performance of our algorithm is encouraging and can be comparable to other algorithms.

In the using of remote sensing technology for agriculture investigation, it is often required to extract plant and vegetation distribution from land covers, especially when plants are sparsely dispersed. Otherwise it would consume considerable expense to perform ground survey. Due to its narrow spectral bandwidth and high spectral resolution, hyperspectral remote sensing technique is regarded as a suitable technique for such task. Conventional classification techniques often suffer from unrealistic mathematical distribution assumption, and are also subject to the lack of prerequisite information. Trying to address the problem of unsurprised plant classification, this paper proposes an Independent Component Analysis (ICA) based feature extraction algorithm. ICA is a technique that stems out from the Blind Source Separation. In hyperspectral data processing, ICA projects the vector to the space where the items of the vector are mutually statistically independent, and therefore is capable of extracting various kinds of information from imagery. Besides, so as to strengthen the contrast of result independent component, a mathematical morphology based post-processing procedure is appended after ICA decomposition. Through real hyperspectral data experiments, our algorithm has demonstrated better performance in classification. Computation efficiency and noise robustness have also been improved by an extra noise filtering effort.

We present a new approach for extending the particle swarm optimization algorithm to multi-optima problems by using ideas from possibility theory. An elastic constraint is used to let the particles dynamically explore the solution space in two phases. In the exploratory phase, particles explore the space in an effort to track the global minima while also traversing the local minima. In the exploitatory phase, particles disperse in the local neighborhoods to locate the best local minima. The proposed PPSO has been applied to data clustering and object detection. Our preliminary results indicate that the proposed approach is efficient and robust.

Generic object detection and recognition systems need to be able to recognize objects even if they occur at arbitrary scales, or shown from different perspectives on highly textured backgrounds. This problem has recently gained a lot of attention in the field of computer vision e.g. Agarwal and Roth [1], Fergus et al. [2] and Opelt et al. [3]. We propose several modifications to the framework of generic object recognition system as described in [3]. At first, we use K-means to cluster the features into a uniform frame in order to obtain a simple feature vector per image. Secondly, we hypothesis that by combining the distinct features using Kernel Canonical Correlation Analysis (KCCA) we would be able to increase the classification power (Vinokourov et al. [4]). Finally, we use a Support Vector Machine (SVM) classifier in the semantic space obtained by KCCA. In our experiments we compare our method to SVM on the raw data and to the results published in [2, 3]. We are able to show that our proposed approach is able to achieve improved performance on both simple [2,3] and difficult [3] datasets. And the overall complexity of our system is significantly lower than that in [3].

In this paper, we apply independent component analysis (ICA) to the reduction of spatially correlated additive noise in images. We take a degraded image as the mixture of the noise and the original image, which are statistically independent. From a view of blind signal separation, we try to restore the original image from two linear mixtures. Motivated by the fact that autocorrelation exists in the neighborhoods of the image and the noise; we design another mixture using the diffusion equation. Then we employ independent component analysis to separate the image and the noise from the two mixtures. Simulation experiments are carried out to remove the Poisson noise from images. Experimental results indicate and impressive performance of the proposed method. Furthermore, the proposed method can be combined with the Wavelet Shrinkage method to improve the denoising performance.

In this paper, we propose a two-stage method for efficient object detection that combines full advantages of AdaBoost and SVM to achieve a reasonable balance in both training and classification speed. In the first stage, we use Haar wavelet features and AdaBoost to train a cascade of classifiers for quick and efficient rejection. This cascade of classifiers consists of simple-to-complex classifiers that allow adapting to complexity of input patterns, rejects almost 90%-99% non-object patterns rapidly. Hard patterns, object-like patterns, which are passed through the first stage, will be classified by the second stage which uses a non linear SVM-based classifier with pixel-based global features. The nonlinear SVM classifier is robust enough in order to reach high performance. We have investigated our proposed method to detect different kinds of objects such as face and facial features like eye and mouth regions. In training, our system is roughly 25 times faster than the system trained by AdaBoost. In running, the experimental results show 1,000 times faster than SVM based method and slightly slower than AdaBoost based method with a comparable accuracy.

In modern warfare, timely extraction of coordinates from tactical images remains a critical restriction in the targeting loop. Often, a warfighter has immediate access to real time sensor images. However, extracting the true coordinates of a target can be time consuming and difficult. Usually, the sensor will provide basic parameters such as range, heading, and depression angle, which can be used to correlate the image with an on-line database of stereo images. However, these parameters are often in error. This paper documents a research project conducted at the Naval Air Warfare Center, China Lake, California which solves this problem for range and heading values within the Digital Precision Strike Suite correlation and targeting software also developed at this facility. Stereo templates are built using the given range or heading and correlated with the tactical image. Three values from the results of the correlation attempt are fed to a neural network which will determine whether the range or heading is too small or too large. This information is used by a binary search algorithm to adjust the given parameter, perform the correlation again and feed the new values to the neural network. The results obtained are very promising.

When a 2D binary image, e.g., the edge-detected boundary of a feature E1 (e.g., the left eye) in a digital image IM1 (e.g., the digital facial picture of a person P1), is allowed to vary between two extreme boundary curves C11 and C12 (e.g., C11 is the boundary curve of the left eye of P1 when he is wearing a serious facial expression, and C12 is that of P1 when he is wearing a beamingly smile expression), we can then select the center of mass (CM) points of the curves and obtain two radial functions to represent these two curves: R11 = f11(theta), R12=f12(theta) where R11, R12, are the radial coordinates of C11, C12 at angle theta with the (polar) coordinates centered at the CM point. Similarly, for a similar IM2 (of a person P2), we can obtain similar boundary curves R21=f21(theta), and R22=f22(theta) (for the left eye of P2 in the two extreme facial expressions). If we sample the analog data R11, R12, R21, R22 at 36 different theta angles with 10-degree increments between adjacent angles, then we will obtain 4 ANALOG vectors of 36-dimensions each. The mapping of the extreme boundaries R11, R12 to +1 (for identifying P1) and R21, R22 to -1 (for identifying P2) can then be used as the learning map required in the design of a neural network for carrying out this task of a precision, automatic identification of two very similar patterns. This paper reports the theory and the experiment of a novel, non-iterative, neural network that will learn this extreme-boundary mapping derived from any two VERY SIMILAR patterns. It will learn not only the static features of each pattern, but also the dynamic variation ranges of the features in each pattern. After the learning, it will then AUTOMATICALLY identify any UNTRAINED pattern (which is ANY gradual change between the two extremes boundaries) as belonging to IM1 or IM2. The neural network DOES NOT have to learn any of these gradual changes one by one.

An intelligent decision support tool to the Radiologist in telemedicine is described. Medical prescriptions are given based on the images of cyst that has been transmitted over computer networks to the remote medical center. The digital image, acquired by sonography, is converted into an intensity image. This image is then subjected to image preprocessing which involves correction methods to eliminate specific artifacts. The image is resized into a 256 x 256 matrix by using bilinear interpolation method. The background area is detected using distinct block operation. The area of the cyst is calculated by removing the background area from the original image. Boundary enhancement and morphological operations are done to remove unrelated pixels. This gives us the cyst volume. This segmented image of the cyst is sent to the remote medical center for analysis by Knowledge based artificial Intelligent Decision Support System (KIDSS). The type of cyst is detected and reported to the control mechanism of KIDSS. Then the inference engine compares this with the knowledge base and gives appropriate medical prescriptions or treatment recommendations by applying reasoning mechanisms at the remote medical center.

The curves and lines in an edged-detected binary image can be analyzed using the adaptive-window detection technique. This window at first moves from the top-left corner of the image frame, and then scans horizontally and downward until it hits the starting point S of a “continuous” line or curve. Then it will automatically track the direction of the curve until it hits an end point E or a branch point B. The coordinates of the starting point S, the end point E or the branch point B will be automatically recorded in a data file, and so are the coordinates of all continuous points between S, E or S, B. For the branch point, the adaptive window will detect how many branches are connected to point B, and it will track automatically each branch until another end point or another branch point is hit. The coordinates of all continuous points between any pair of (cusp) points S, B; S, E; B, B; or B, E, will be automatically recorded in a different data file. Each data file then represents a single curve between 2 cusp points. These data file can then be used to find the analytical expression for each curve. We use polynomial, least square curve fitting techniques to get a very compact set of analytical data for representing or reconstructing the original binary image. This paper reports the image-processing steps, the programming algorithm, and the experimental results on this novel feature extraction technique. It will be verified in each experiment by the reconstruction of the original image from the compactly extracted analog data lines. These data lines can then be used very efficiently for inputting to a neural network for an accurate and yet robust pattern recognition job.

Most current on-line signature verification systems are pen-based, i.e., the signature is written by a pen on a digital tablet. To apply signature verification techniques to Internet applications, one desirable way is to use mouse as the input device. In this paper, a mouse based on-line signature verification system is presented. The signature data is represented in a compact and secure way. The enrollment is convenient for users and the verification can be real-time. The system architecture is a simple client-server structure that suits the common internet applications. The client side posts a sequence of X*Y to the server. That is, the raw signature sequence [X,Y] is represented by X+Y(dot sum of the X-, Y-coordinate sequences). Experimental results show that X+Yis effective for verification. Even if the sequence X+Y is intercepted during transmission over the Internet, original shape of signature [X, Y] can be recovered. To make the system easy to implement, the client side does nothing complex except posting X+Yto the server. The server side preprocesses the sequence and saves it to database during enrollment or match it against the claimed template during verification. The matching technique adopted here is the similarity measure ER². Simulation results show that the mouse based verification system is reliable and secure for internet applications.