Automatic target recognition using forward-looking infrared imagery is a challenging problem because of the highly unpredictable nature of target thermal signatures. The high variability of target signatures, target obscuration, and clutter in the background results in distortion of target features, which are used by the target detection stage to identify a potential target. Consequently, the target detection stage produces a large number of false alarms. Distorted features in the potential targets also make accurate classification of targets difficult. The proposed technique, in essence attempts to repair the distorted features of the targets to improve the target detection/classification accuracy. The proposed technique completes the feature extraction process in two steps: First, the feature vectors are extracted and classified either as complete or incomplete features using feed-forward neural networks. The incomplete features are then transformed into complete features. These features can then be used to identify/classify the targets.

Object detection is an enabling technology that plays a key role in many application areas, such as content based media retrieval. Attentive cognitive vision systems are here proposed where the focus of attention is directed towards the most relevant target. The most promising information is interpreted in a sequential process that dynamically makes use of knowledge and that enables spatial reasoning on the local object information. The presented work proposes an innovative application of attention mechanisms for object detection which is most general in its understanding of information and action selection. The attentive detection system uses a cascade of increasingly complex classifiers for the stepwise identification of regions of interest (ROIs) and recursively refined object hypotheses. While the most coarse classifiers are used to determine first approximations on a region of interest in the input image, more complex classifiers are used for more refined ROIs to give more confident estimates. Objects are modelled by local appearance based representations and in terms of posterior distributions of the object samples in eigenspace. The discrimination function to discern between objects is modeled by a radial basis functions (RBF) network that has been compared with alternative networks and been proved consistent and superior to other artifical neural networks for appearance based object recognition. The experiments were led for the automatic detection of brand objects in Formula One broadcasts within the European Commission's cognitive vision project DETECT.

This paper presents a new approach for extracting spatial features of images based on the theory of regionalized variables. These features can be effectively used for automatic recognition of logo images using neural networks. Experimental results on a public-domain logo database show the effectiveness of the proposed approach.

In this paper a learning algorithm of synergetic neural network based on selective attention parameters is proposed. According to the mechanism of the Human Visual System (HVS), the weight matrix of synergetic neural network can be obtained by multiplying the prototype matrix by selective attention parameters. Two selective attention models based on the human visual system are put forward in this paper. The comparative experiments between the traditional algorithm SCAP and the new method we proposed in the application of recognizing the real gray images of numeric and alphabetic characters are done. And the results show that our method can improve the synergetic neural network's recognition performance and be more suitable to human visual system.

This paper describes a cost-effective, real-time (640x480 at 30Hz) upright frontal face detector as part of an ongoing project to develop a video-based, tetherless 3D head position and orientation tracking system. The work is specifically targeted for auto-stereoscopic displays and projection-based virtual reality systems. The proposed face detector is based on a modified LAMSTAR neural network system. At the input stage, after achieving image normalization and equalization, a sub-window analyzes facial features using a neural network. The sub-window is segmented, and each part is fed to a neural network layer consisting of a Kohonen Self-Organizing Map (SOM). The output of the SOM neural networks are interconnected and related by correlation-links, and can hence determine the presence of a face with enough redundancy to provide a high detection rate. To avoid tracking multiple faces simultaneously, the system is initially trained to track only the face centered in a box superimposed on the display. The system is also rotationally and size invariant to a certain degree.

A method for recognizing faces in relativley unconstrained environments, such as offices, is described. It can recognize faces occurring over an extended range of orientations and distances relative to the camera. As the pattern recognition mechanism, a bank of small neural networks of the multilayer perceptron type is used, where each perceptron has the task of recognizing only a single person's face. The perceptrons are trained with a set of nine face images representing the nine main facial orientations of the person to be identified, and a set face images from various other persons. The center of the neck is determined as the reference point for face position unification. Geometric normalization and reference point determination utilizes 3-D data point measurements obtained with a stereo camera. The system achieves a recognition rate of about 95%.

The hybrid evolutionary algorithm is used for image registration formulated as an optimization problem of finding a vector of parameters minimizing the difference between images. The reproduction phase of the algorithm is enhanced with a two-level operation of local correction performed on the best genes in the reproduction pool. Random search is performed in the neighborhood of a gene until the time interval reaches a pre-set threshold. If the gene still retains its position in the pool, a refined multi-step search is performed using the Downhill simplex method. In order to improve the computational performance of the local search, local response analysis is used in the following way. All domains of the given reference image are classified according to their local response to a unit variation of the parameter vector. The classification scheme is based on a self-organizing neural network. During the local correction of the reproduction pool, the step size in the Downhill simplex search is modified according to the class of the image domain.

Systems for processing high resolution images need to be fast, compact, and efficient. Image processing systems that incorporate optics into its architecture can provide the speed and potentially the compactness to meet the demands of analyzing images. In this paper a hybrid approach to image analysis using Winner Take All neural network dynamics with optical and electronic implementation is discussed. Resulting images from the system simulations are explored for use in object and background discrimination for image segmentation tasks.

In this paper, we discuss to develop automatic classification system for true color Leukocyte image. In view of the deficiencies of traditional combination optimization method, a new method based on genetic algorithm is proposed. Combining the specific situation of cell classification, we made some modification. Finally neural network with error back-propagation is training using the selected feature sets. The result shows this method optimize the classification performance.

In this paper, a stereo matching approach for 3D reconstruction based on wavelet analysis is presented. It can be used in neuro-vision system. The approach can be divided into two parts. First, the stereo matching problem is solved with wavelet analysis. Dyadic discrete wavelet analysis is adopted in this process and stereo matching process is realized with global optimization. A coherent hierarchical matching strategy is constructed, so that the stereo matching process can be accomplished with coarse to fine techniques. Second, a 3D object reconstruction neural network is constructed by using BP neural network. By feeding the image corresponding points between the left image and right image in a stereo image pair, the 3D coordinates of points on object surface can be obtained using this neural network and the configuration and shape of the object can be reconstructed. With multiple 3D reconstruction neural networks the 3D reconstruction processes can be performed in parallel. The examples for both synthetic and real images are shown in the experiment, and good results are obtained.

This paper presents a 3-D reconstruction method IBM (image based modeling) of an image that does not contain any camera information. This system adopts a 3D reconstruction method based on a model. Model-based 3D reconstruction recovers an image using the geometric characteristics of a pre-defined polyhedron model. It uses a pre-defined polyhedron model as the primitive and the 3D reconstruction is processed by mapping the correspondence point of the primitive model onto the picture image. Existing model-based 3D reconstruction methods were used for the reconstruction of camera parameters or error method through iteration. However, we proposed a method for a primitive model that uses the segment and the center of the segment for the reconstruction process. This method enables the reconstruction of the primitive model to be processed using the minimum camera parameters (e.g. focal length) during the segment reconstruction process.

Electrical impedance tomography (EIT) is a new computer tomography technology, which reconstructs an impedance (resistivity, conductivity) distribution or change by making voltage and current measurements on the object’s periphery. Image reconstruction in EIT is an ill-posed, non-linear inverse problem. A method deciding the place of impedance change for EIT is proposed in this paper, in which a multilevel BP neural network (MBPNN) is used to express the non-linear relation between the impedance change inside the object and the voltage change measured on the surface of the object. Thus, the location of the impedance change can be decided by the measured voltage variation on the surface, and then the impedance change will be reconstructed with linear approximated method. MBPNN can decide the impedance change location exactly without needing long training time. It alleviates some noise affection and can be expanded, which makes sure about the high precision and space resolution of the reconstructed image that can’t be accessed by the back projection method.

A feedback neural network (FBNN) can be triggered by ANY input analog pattern vector. Then depending on the domain-of-convergence (or domain-of-attraction in the languages of nonlinear systems) that this triggering pattern falls into, the FBNN will go around and around the feedback loop and finally settle down at one of the few designated patterns associatively stored in the connection matrix. This recalled (or the settle-down) pattern will stay at the output even when the input triggering pattern is removed because of the self-sustained feedback action of the FBNN. The triggering pattern does not have to be the same as the stored pattern that it recalls. It can be a noise-affected pattern. But as long as it falls within the designated noise range (or the designated domain of convergence) of an accurately stored pattern, that accurate pattern will be recalled and permanently appear at the output even when the input triggering is removed. This paper derives, from the principle of NONITERATIVE LEARNING, the basic design method of this FBNN with controlled domains-of-convergence taken into account in the design.

A neural network based image enhancement method is introduced to improve the image resolution from a sequence of low resolution image frames. Most of the existing methods reconstruct a high-resolution image from a multiple of low-resolution image frames by minimizing some established cost function using a mathematical technique. This method, however, uses an integrated recurrent neural network (IRNN) that is particularly designed to be capable of learning an optimal mapping from a multiple of low-resolution image frames to a high-resolution image through training. The IRNN consists of four feed-forward sub-networks working collectively with the ability of having a feedback of information from its output to input. As such, it is capable of both learning and searching the optimal solution in the solution space leading to high resolution images. Simulation results demonstrate that the proposed IRNN has good potential in solving image resolution enhancement problem, as it can adapt itself to the various conditions of the reconstruction problem by learning.