This PDF file contains the front matter associated with SPIE Proceedings Volume 9844, including the Title Page, Copyright information, Table of Contents, Introduction (if any), and Conference Committee listing.

The problem this paper addresses is to create an approximation for a source image given only a randomly selected subset of pixel samples extracted from a blurred version of the source image. This problem is different from the conventional image restoration problem, which attempts to create an approximation for the source image given all of the pixel samples available in the blurred image. Our approach finds a minimum weighted L2 norm solution for the ideal image that satisfies linear constraints given by the observed samples of the blurred image.

The Photo-G program conducted by Naval Air Systems Command at the Atlantic Test Range in Patuxent River, Maryland, uses photogrammetric analysis of large amounts of real-world imagery to characterize the motion of objects in a 3-D scene. Current approaches involve several independent processes including target acquisition, target identification, 2-D tracking of image features, and 3-D kinematic state estimation. Each process has its own inherent complications and corresponding degrees of both human intervention and computational complexity. One approach being explored for automated target acquisition relies on exploiting the pixel intensity distributions of photogrammetric targets, which tend to be patterns with bimodal intensity distributions. The bimodal distribution partitioning algorithm utilizes this distribution to automatically deconstruct a video frame into regions of interest (ROI) that are merged and expanded to target boundaries, from which ROI centroids are extracted to mark target acquisition points. This process has proved to be scale, position and orientation invariant, as well as fairly insensitive to global uniform intensity disparities.

Automatic target detection in infrared scenes is a vital task for many application areas like defense, security and border surveillance. For anti-ship missiles, having a fast and robust ship detection algorithm is crucial for overall system performance. In this paper, a straight-forward yet effective ship detection method for infrared scenes is introduced. First, morphological grayscale reconstruction is applied to the input image, followed by an automatic thresholding onto the suppressed image. For the segmentation step, connected component analysis is employed to obtain target candidate regions. At this point, it can be realized that the detection is defenseless to outliers like small objects with relatively high intensity values or the clouds. To deal with this drawback, a post-processing stage is introduced. For the post-processing stage, two different methods are used. First, noisy detection results are rejected with respect to target size. Second, the waterline is detected by using Hough transform and the detection results that are located above the waterline with a small margin are rejected. After post-processing stage, there are still undesired holes remaining, which cause to detect one object as multi objects or not to detect an object as a whole. To improve the detection performance, another automatic thresholding is implemented only to target candidate regions. Finally, two detection results are fused and post-processing stage is repeated to obtain final detection result. The performance of overall methodology is tested with real world infrared test data.

Pre-processing, thresholding and post-processing stages are very important especially for very small target detection from infrared images. The effects of these stages to the final detection performance are measured in this study. Various methods for each stage are compared based on the final detection performance, which is defined by precision and recall values. Among various methods, the best method for each stage is selected and proved. For the pre-processing stage, local block based methods perform the best, nearly for all thresholding methods. The best thresholding method is chosen as the one, which does not need any user defined parameter. Finally, the post processing method, which is suitable for the best performing pre-processing and thesholding methods is selected.

Sum of square difference (SSD) and normalized cross correlation (NCC) are two different template matching techniques and their fast implementations have been investigated independently. The SSD approach is known to be simple and fast, however it is variant to image intensity change that lead to low performance. On the other hand, the NCC method is invariant to intensity change and has high performance, but its computational cost is high. In this paper, we derive an equation that connects NCC and SSD. From this equation, we propose SSD based partial elimination for the fast implementation of NCC template matching. This new technique takes the advantages of both NCC’s high performance and SSD’s low computational cost. It is fast and has high performance. Then we propose a uniform smoothing approach that further reduces computational cost for NCC. Experiments show that the proposed method is significantly faster than the techniques reported in literature.

Publisher’s Note: This paper, originally published on 5/12/2016, was replaced with a corrected/revised version on 5/18/2016. If you downloaded the original PDF but are unable to access the revision, please contact SPIE Digital Library Customer Service for assistance.

When dealing with sparse or no labeled data in the target domain, transfer learning shows its appealing performance by borrowing the supervised knowledge from external domains. Recently deep structure learning has been exploited in transfer learning due to its attractive power in extracting effective knowledge through multi-layer strategy, so that deep transfer learning is promising to address the cross-domain mismatch. In general, cross-domain disparity can be resulted from the difference between source and target distributions or different modalities, e.g., Midwave IR (MWIR) and Longwave IR (LWIR). In this paper, we propose a Weighted Deep Transfer Learning framework for automatic target classification through a task-driven fashion. Specifically, deep features and classifier parameters are obtained simultaneously for optimal classification performance. In this way, the proposed deep structures can extract more effective features with the guidance of the classifier performance; on the other hand, the classifier performance is further improved since it is optimized on more discriminative features. Furthermore, we build a weighted scheme to couple source and target output by assigning pseudo labels to target data, therefore we can transfer knowledge from source (i.e., MWIR) to target (i.e., LWIR). Experimental results on real databases demonstrate the superiority of the proposed algorithm by comparing with others.

Algorithms to automatically recognize ship type from satellite imagery are desired for numerous maritime applications. This task is difficult, and example imagery accurately labeled with ship type is hard to obtain. Convolutional neural networks (CNNs) have shown promise in image recognition settings, but many of these applications rely on the availability of thousands of example images for training. This work attempts to under- stand for which types of ship recognition tasks CNNs might be well suited. We report the results of baseline experiments applying a CNN to several ship type classification tasks, and discuss many of the considerations that must be made in approaching this problem.

As part of a broader research effort in multispectral image analysis, an improved segmentation algorithm based on the classical Watershed concept was developed. A requirement for this research was to develop a segmentation algorithm that could effectively extract objects of interest in both visual and thermal image pairs. The classical Watershed algorithm can be enhanced with "markers" identifying clusters of pixels belonging to the same object or to the background. There are several ways to create the markers and the proposed Watershed with Thermal Markers allows the user to extract objects of interest from both visual and/or thermal dataset using an initial seed extracted from the thermal image.

This research demonstrates how inexpensive commercial off-the-shelf lighting components and microcontrollers can be used to construct a solution for occupant and asset localization and tracking through visible light communication (VLC). Through the modulation of the emitted light from networked LED luminaires, the location of a receiver can be determined. This paper describes the implementation of the VLC enabled LED luminaires, in addition to the infrared synchronization protocol, which enabled inexpensive white LEDs to be time division multiplexed to avoid packet collisions. Luminaires use token message passing to regulate packet transmission. Physical construction of these luminaires is discussed in addition to the simulated performance of this system.

Natural and human-induced disturbances are resulting in degradation and loss of seagrass. Freshwater flooding, severe meteorological events and invasive species are among the major natural disturbances. Human-induced disturbances are mainly due to boat propeller scars in the shallow seagrass meadows and anchor scars in the deeper areas. Therefore, there is a vital need to map seagrass ecosystems in order to determine worldwide abundance and distribution. Currently there is no established method for mapping the pothole or scars in seagrass. One of the most precise sensors to map the seagrass disturbance is side scan sonar. Here we propose an automatic method which detects seagrass potholes in sonar images. Side scan sonar images are notorious for having speckle noise and uneven illumination across the image. Moreover, disturbance presents complex patterns where most segmentation techniques will fail. In this paper, by applying mathematical morphology technique and calculating the local standard deviation of the image, the images were enhanced and the pothole patterns were identified. The proposed method was applied on sonar images taken from Laguna Madre in Texas. Experimental results show the effectiveness of the proposed method.

Video anomaly detection is a critical research area in computer vision. It is a natural first step before applying object recognition algorithms. There are many algorithms that detect anomalies (outliers) in videos and images that have been introduced in recent years. However, these algorithms behave and perform differently based on differences in domains and tasks to which they are subjected. In order to better understand the strengths and weaknesses of outlier algorithms and their applicability in a particular domain/task of interest, it is important to measure and quantify their performance using appropriate evaluation metrics. There are many evaluation metrics that have been used in the literature such as precision curves, precision-recall curves, and receiver operating characteristic (ROC) curves. In order to construct these different metrics, it is also important to choose an appropriate evaluation scheme that decides when a proposed detection is considered a true or a false detection. Choosing the right evaluation metric and the right scheme is very critical since the choice can introduce positive or negative bias in the measuring criterion and may favor (or work against) a particular algorithm or task. In this paper, we review evaluation metrics and popular evaluation schemes that are used to measure the performance of anomaly detection algorithms on videos and imagery with one or more anomalies. We analyze the biases introduced by these by measuring the performance of an existing anomaly detection algorithm.

We investigate the application of learned convolutional filters in multi-object visual tracking. The filters were learned in both a supervised and unsupervised manner from image data using artificial neural networks. This work follows recent results in the field of machine learning that demonstrate the use learned filters for enhanced object detection and classification. Here we employ a track-before-detect approach to multi-object tracking, where tracking guides the detection process. The object detection provides a probabilistic input image calculated by selecting from features obtained using banks of generative or discriminative learned filters. We present a systematic evaluation of these convolutional filters using a real-world data set that examines their performance as generic object detectors.

As Unmanned Aerial Systems grow in numbers, pedestrian detection from aerial platforms is becoming a topic of increasing importance. By providing greater contextual information and a reduced potential for occlusion, the aerial vantage point provided by Unmanned Aerial Systems is highly advantageous for many surveillance applications, such as target detection, tracking, and action recognition. However, due to the greater distance between the camera and scene, targets of interest in aerial imagery are generally smaller and have less detail. Deep Convolutional Neural Networks (CNN’s) have demonstrated excellent object classification performance and in this paper we adopt them to the problem of pedestrian detection from aerial platforms. We train a CNN with five layers consisting of three convolution-pooling layers and two fully connected layers. We also address the computational inefficiencies of the sliding window method for object detection. In the sliding window configuration, a very large number of candidate patches are generated from each frame, while only a small number of them contain pedestrians. We utilize the Edge Box object proposal generation method to screen candidate patches based on an "objectness" criterion, so that only regions that are likely to contain objects are processed. This method significantly reduces the number of image patches processed by the neural network and makes our classification method very efficient. The resulting two-stage system is a good candidate for real-time implementation onboard modern aerial vehicles. Furthermore, testing on three datasets confirmed that our system offers high detection accuracy for terrestrial pedestrian detection in aerial imagery.

We consider the problem of recognizing a particular target of interest (i.e. the "correct" target) while rejecting other targets and background clutter. In such instances, the probability of recognizing the correct target (P_CT) is a suitable metric for assessing the performance of the target recognition algorithm. We present a definition for PCT and illustrate how it differs from conventional metrics for target recognition by means of an example. It is further shown that an adaptive target recognition algorithm, which relies on track position to obtain multiple looks at the target, can significantly improve P_CT while reducing the track uncertainty.

In this paper a novel method is described for representation and classification of target by random graphs. A target is represented in terms of set primitives that jointly represent a random graph structure. Random graph is a graph structure with randomly varying vertex and arc attribute values. Random graphs and their statistical and matrix representations are useful when one encounters the problem of classifying signatures of partially occluded targets. We present a number of observations that spectra of random graphs of partially occluded and non-occluded target signatures are related through an interlacing rule and the correlations of their Laplacians lead to robust classification.

We consider the transformation of the Raleigh distribution into a new distribution so that the new distribution behaves approximately the same as the Rayleigh for small values of the argument but becomes heavy tailed for large values.

We apply the continuous wavelet transform to dispersive pulse propagation and obtain an approximation that is easily applied. The approximation shows that one can evolve the wavelet transform of the pulse in a simple manner, by calculating the wavelet transform at time zero and making a simple algebraic substitution.

Radar target identification using probabilistic vector machines is investigated and tested using real radar data collected in a compact range for commercial aircraft models. Unlike relevance vector machines (RVM) that utilize zero-mean Gaussian prior for every weight for both negative and positive classes and are thus vulnerable to questionable (deceptive) vectors, probabilistic vector machines [2], alternatively, use nonnegative priors for the positive class and vice versa. This paper compares the performance of these machines with other target identification tools, and highlights scenarios where classification via a probabilistic vector machine is more plausible. The problem addressed in this paper is a M-ary target classification problem and is implemented as a set of pairwise comparisons between all competing hypotheses.

Supervised multi-class target recognition algorithms label an input pattern according to the most similar training class. Typically, the number of training classes is fixed and known a priori. In practice, however, a classifier may encounter novel targets that were not seen in training and label them incorrectly. Recent work in open set recognition (OSR) develops classifiers that can identify training targets as well as previously unknown targets. This results in a reduced number of forced misclassifications by "ejecting" targets that were not present in training. Several OSR algorithms are based on support vector machines (SVMs), namely, the 1-vs-set machine, W-SVM, and POS-SVM. The 1-vs-set machine, a linear classifier, forms a "lab" around each training class to discriminate it from the remaining training classes and limit the risk of labeling open space as target space. The W-SVM uses a novel dual-calibration technique to map the SVM outputs to posterior probabilities, which are then subjected to a pair of user-specified thresholds. The POS-SVM relies on a single calibration step, but features data-driven posterior probability thresholds that are chosen automatically. Both the W-SVM and POS-SVM have the capability to use nonlinear SVM kernel functions and perform particularly well with the popular Gaussian RBF kernel. Past works have shown that these algorithms can be effective for classifying ladar and IR images with a rejection option. In this paper, we apply these algorithms to the MSTAR SAR dataset and analyze their performance for classifying known targets and rejecting unknown targets in the presence of clutter.

Computer and machine vision applications are used in numerous fields to analyze static and dynamic imagery in order to assist or automate decision-making processes. Advancements in sensor technologies now make it possible to capture and visualize imagery at various wavelengths (or bands) of the electromagnetic spectrum. Multispectral imaging has countless applications in various fields including (but not limited to) security, defense, space, medical, manufacturing and archeology. The development of advanced algorithms to process and extract salient information from the imagery is a critical component of the overall system performance.

The fundamental objective of this research project was to investigate the benefits of combining imagery from the visual and thermal bands of the electromagnetic spectrum to improve the recognition rates and accuracy of commonly found objects in an office setting. A multispectral dataset (visual and thermal) was captured and features from the visual and thermal images were extracted and used to train support vector machine (SVM) classifiers. The SVM’s class prediction ability was evaluated separately on the visual, thermal and multispectral testing datasets.

Dimension reduction techniques have become one popular unsupervised approach used towards detecting anomalies in hyperspectral imagery. Although demonstrating promising results in the literature on specific images, these methods can become difficult to directly interpret and often require tuning of their parameters to achieve high performance on a specific set of images. This lack of generality is also compounded by the need to remove noise and atmospheric absorption spectral bands from the image prior to detection. Without a process for this band selection and to make the methods adaptable to different image compositions, performance becomes difficult to maintain across a wider variety of images. Here, we present a framework that uses factor analysis to provide a robust band selection and more meaningful dimension reduction with which to detect anomalies in the imagery. Measurable characteristics of the image are used to create an automated decision process that allows the algorithm to adjust to a particular image, while maintaining high detection performance. The framework and its algorithms are detailed, and results are shown for forest, desert, sea, rural, urban, anomaly-sparse, and anomaly-dense imagery types from different sensors. Additionally, the method is compared to current state-of-the-art methods and is shown to be computationally efficient.

Visual tracking is important in computer vision. At present, although many algorithms of visual tracking have been proposed, there are still many problems which are needed to be solved, such as occlusion and frame speed. To solve these problems, this paper proposes a novel method which based on compressive tracking. Firstly, we make sure the occlusion happens if the testing result about image features by the classifiers is lower than a threshold value which is certain. Secondly, we mark the occluded image and record the occlusion region. In the next frame, we test both the classifier and the marked image. This algorithm makes sure the tracking is fast, and the result about solving occlusion is much better than other algorithms, especially compressive tracking.

In this work, array processing techniques based on subspace decomposition of signal have been evaluated for estimation of direction of arrival of moving targets using acoustic signatures. Three subspace based approaches – Incoherent Wideband Multiple Signal Classification (IWM), Least Square-Estimation of Signal Parameters via Rotation Invariance Techniques (LS-ESPRIT) and Total Least Square- ESPIRIT (TLS-ESPRIT) are considered. Their performance is compared with conventional time delay estimation (TDE) approaches such as Generalized Cross Correlation (GCC) and Average Square Difference Function (ASDF). Performance evaluation has been conducted on experimentally generated data consisting of acoustic signatures of four different types of civilian vehicles moving in defined geometrical trajectories. Mean absolute error and standard deviation of the DOA estimates w.r.t. ground truth are used as performance evaluation metrics. Lower statistical values of mean error confirm the superiority of subspace based approaches over TDE based techniques. Amongst the compared methods, LS-ESPRIT indicated better performance.

The use of aerial hyperspectral imagery for the purpose of remote sensing is a rapidly growing research area. Currently, targets are generally detected by looking for distinct spectral features of the objects under surveillance. For example, a camouflaged vehicle, deliberately designed to blend into background trees and grass in the visible spectrum, can be revealed using spectral features in the near-infrared spectrum. This work aims to develop improved target detection methods, using a two-stage approach, firstly by development of a physics-based atmospheric correction algorithm to convert radiance into re ectance hyperspectral image data and secondly by use of improved outlier detection techniques. In this paper the use of the Percentage Occupancy Hit or Miss Transform is explored to provide an automated method for target detection in aerial hyperspectral imagery.

We have developed a new way for detection and tracking of human full-body and body-parts with color (intensity) patch morphological segmentation and adaptive thresholding for security surveillance cameras. An adaptive threshold scheme has been developed for dealing with body size changes, illumination condition changes, and cross camera parameter changes. Tests with the PETS 2009 and 2014 datasets show that we can obtain high probability of detection and low probability of false alarm for full-body. Test results indicate that our human full-body detection method can considerably outperform the current state-of-the-art methods in both detection performance and computational complexity. Furthermore, in this paper, we have developed several methods using color features for detection and tracking of human body-parts (arms, legs, torso, and head, etc.). For example, we have developed a human skin color sub-patch segmentation algorithm by first conducting a RGB to YIQ transformation and then applying a Subtractive I/Q image Fusion with morphological operations. With this method, we can reliably detect and track human skin color related body-parts such as face, neck, arms, and legs. Reliable body-parts (e.g. head) detection allows us to continuously track the individual person even in the case that multiple closely spaced persons are merged. Accordingly, we have developed a new algorithm to split a merged detection blob back to individual detections based on the detected head positions. Detected body-parts also allow us to extract important local constellation features of the body-parts positions and angles related to the full-body. These features are useful for human walking gait pattern recognition and human pose (e.g. standing or falling down) estimation for potential abnormal behavior and accidental event detection, as evidenced with our experimental tests. Furthermore, based on the reliable head (face) tacking, we have applied a super-resolution algorithm to enhance the face resolution for improved human face recognition performance.

Human activity recognition has often prioritized low-level features extracted from imagery or video over higher-level class attributes and ontologies because they have traditionally been more effective on small datasets. However, by including knowledge-driven associations between actions and attributes while recognizing the lower-level attributes with their temporal relationships, we can attempt a hybrid approach that is more easily extensible to much larger datasets. We demonstrate a combination of hard and soft features with a comparison factor that prioritizes one approach over the other with a relative weight. We then exhaustively search over the comparison factor to evaluate the performance of a hybrid human activity recognition approach in comparison to the base hard approach at 84% accuracy and the current state-of-the-art.

In this work, the data covariance matrix is diagonalized to provide an orthogonal bases set using the eigen vectors of the data. The eigen-vector decomposition of the data is transformed and filtered in the transform domain to truncate the data for robust features related to a specified set of targets. These truncated eigen features are then combined and reconstructed to utilize in a composite filter and consequently utilized for the automatic target detection of the same class of targets. The results associated with the testing of the current technique are evaluated using the peak-correlation and peak-correlation energy metrics and are presented in this work. The inverse transformed eigen-bases of the current technique may be thought of as an injected sparsity to minimize data in representing the skeletal data structure information associated with the set of targets under consideration.

Origami devices have the ability to spatially reconfigure between 2D and 3D states through folding motions. The precise mapping of origami presents a novel method to spatially tune radio frequency (RF) devices, including adaptive antennas, sensors, reflectors, and frequency selective surfaces (FSSs). While conventional RF FSSs are designed based upon a planar distribution of conductive elements, this leaves the large design space of the out of plane dimension underutilized. We investigated this design regime through the computational study of four FSS origami tessellations with conductive dipoles. The dipole patterns showed increased resonance shift with decreased separation distances, with the separation in the direction orthogonal to the dipole orientations having a more significant effect. The coupling mechanisms between dipole neighbours were evaluated by comparing surface charge densities, which revealed the gain and loss of coupling as the dipoles moved in and out of alignment via folding. Collectively, these results provide a basis of origami FSS designs for experimental study and motivates the development of computational tools to systematically predict optimal fold patterns for targeted frequency response and directionality.

A novel concept utilizing a hot-carrier effect based on carrier interactions is achieved to extend the wavelength of the photodetector’s spectral response. A detector with a designed wavelength threshold (λ^t) at 3.1 μm displays two different extended thresholds at different temperatures. A very-long wavelength infrared response up to 55 μm was observed up to 35 K; while a threshold wavelength of 8.9 μm was observed in the temperature range 60K − 90K. Response tuning is implemented via varying the degree of hot-hole injection.

We empirically test the capacity of an improved system to identify not just images of individual guns, but partially occluded guns and their parts appearing in a videoframe. This approach combines low-level geometrical information gleaned from the visual images and high-level semantic information stored in an ontology enriched with meronymic part-whole relations. The main improvements of the system are handling occlusion, new algorithms, and an emerging meronomy. Well-known and commonly deployed in ontologies, actual meronomies need to be engineered and populated with unique solutions. Here, this includes adjacency of weapon parts and essentiality of parts to the threat of and the diagnosticity for a weapon. In this study video sequences are processed frame by frame. The extraction method separates colors and removes the background. Then image subtraction of the next frame determines moving targets, before morphological closing is applied to the current frame in order to clean up noise and fill gaps. Next, the method calculates for each object the boundary coordinates and uses them to create a finite numerical sequence as a descriptor. Parts identification is done by cyclic sequence alignment and matching against the nodes of the weapons ontology. From the identified parts, the most-likely weapon will be determined by using the weapon ontology.

Matching facial images acquired in different electromagnetic spectral bands remains a challenge. An example of this type of comparison is matching active or passive infrared (IR) against a gallery of visible face images. When combined with cross-distance, this problem becomes even more challenging due to deteriorated quality of the IR data. As an example, we consider a scenario where visible light images are acquired at a short standoff distance while IR images are long range data. To address the difference in image quality due to atmospheric and camera effects, typical degrading factors observed in long range data, we propose two approaches that allow to coordinate image quality of visible and IR face images. The first approach involves Gaussian-based smoothing functions applied to images acquired at a short distance (visible light images in the case we analyze). The second approach involves denoising and enhancement applied to low quality IR face images. A quality measure tool called Adaptive Sharpness Measure is utilized as guidance for the quality parity process, which is an improvement of the famous Tenengrad method. For recognition algorithm, a composite operator combining Gabor filters, Local Binary Patterns (LBP), generalized LBP and Weber Local Descriptor (WLD) is used. The composite operator encodes both magnitude and phase responses of the Gabor filters. The combining of LBP and WLD utilizes both the orientation and intensity information of edges. Different IR bands, short-wave infrared (SWIR) and near-infrared (NIR), and different long standoff distances are considered. The experimental results show that in all cases the proposed technique of image quality parity (both approaches) benefits the final recognition performance.

We address the problem of determining the source location of an electromagnetic signal from the signal received by one or more moving receivers. We base our process on cross-spectral methods that were developed in the early 1980’s for analysis and demodulation/despreading of communication and spread spectrum signals and were later applied to speech processing and speech enhancement. In this article, we expand the concept of robust polynomial tracking, which we demonstrate may be used to solve for the emitter location in closed form. This is accomplished by generating and solving a system of equations representing curves, each of which passes through the emitter location.

We describe here the design and implementation of a software module that provides both auditory and visual feedback of the eye position measured by a commercially available eye tracking system. The present audio-visual feedback module (AVFM) serves as an extension to the Arrington Research ViewPoint EyeTracker, but it can be easily modified for use with other similar systems. Two modes of audio feedback and one mode of visual feedback are provided in reference to a circular area-of-interest (AOI). Auditory feedback can be either a click tone emitted when the user’s gaze point enters or leaves the AOI, or a sinusoidal waveform with frequency inversely proportional to the distance from the gaze point to the center of the AOI. Visual feedback is in the form of a small circular light patch that is presented whenever the gaze-point is within the AOI. The AVFM processes data that are sent to a dynamic-link library by the EyeTracker. The AVFM’s multithreaded implementation also allows real-time data collection (1 kHz sampling rate) and graphics processing that allow display of the current/past gaze-points as well as the AOI. The feedback provided by the AVFM described here has applications in military target acquisition and personnel training, as well as in visual experimentation, clinical research, marketing research, and sports training.

This paper describes an approach to vessel classification from satellite images using content based image retrieval methodology. Content-based image retrieval is an important problem in both medical imaging and surveillance applications. In many cases the archived reference database is not fully structured, thus making content-based image retrieval a challenging problem. In addition, in surveillance applications, the query image may be affected by weather or/and geometric distortions. Our approach of content-based vessel image retrieval consists of two phases. First, we create a structured reference database, then for each new query image of a vessel we find the closest cluster of images in the structured reference database, thus identifying and classifying the vessel. Then we update the closest cluster with new query image.