Anomaly and Target Detection

Hyperspectral target detection, SAR data analysis, image processing
Researcher: Stanley Rotman
Department: Electrical and Computer Engineering
Faculty: Engineering
E-mail:srotman@ee.bgu.ac.il

My basic research is in the field of target acquisition, whether large targets or sub-pixel targets. My methods for target acquisition are through multi-dimensional data analysis.

Theoretical and applied computational vision, visual psychophysics, human perception, computational neuroscience of the visual system

Researcher: Ohad Ben-Shahar
Department: Computer Science
Faculty: Natural Sciences
E-mail: ben-shahar@cs.bgu.ac.il

HLS is an interdisciplinary domain of application that requires the incorporation of numerous technologies. In large parts, it relies and requires the acquisition of visual (or pseudo visual) information and its automatic processing. Computer vision offers important tools towards both of these goals, and facilitates various relevant tasks such as target detection, surveillance, forensic science, early threat detection, visual data mining, motion analysis and tracking, remote sensing, etc...

Although in my research group we conduct primarily basic research, much of our research program can be useful for HLS applications. In particular, we study

1. The recovering of object shapes from images only – which could be used for HLS applications such as automatic inspection of goods and luggage, the detection of suspected targets in unknown environments, or intelligence gathering from remote sensing devices.

2. The completion of missing visual information – which could assist in the analysis and interpretation of visual scenes with occlusions.

3. Real time tracking in videos – which could be used for surveillance, reconnaissance, and targeting applications.

4. Scene classification and the processing of saliency in images – which can be used for alerts regarding suspected objects or the detection of targets whose appearance is unpredictable.

The use of remote sensing for anomaly and target detection

Researcher: Dan G. Blumberg and Stanley Rotman
Department: Geography and Environmental Development and the Department of Electrical and Computer Engineering
Faculty: Humanities and Social Sciences and Engineering
E-mail: blumberg@bgu.ac.il ; srotman@ee.bgu.ac.il

The use of space and air borne imagery provide an effective way to scan large areas very rapidly. The problem however, with such imagery can be two-fold. First, the immense volume of digital data being gathered can be too large for operators to scan manually. Second, the ability to disguise a hostile object and blend it in to the background. Within the Earth and Planetary Image Facility through collaboration between engineering and geography faculty and students we have improved and developed methods for the detection of targets and anomalies in hyper, multi spectral and radar images. These are done through the use of signal processing techniques with a strong understanding of the natural environment and its appearance in remotely acquired imagery. The use of these automated methods can provide an effective method for the protection of land and marine boundaries and the scanning of vast areas.

Extraction of moving objects from video and image compression

Researcher: Ofer Hadar
Department: Communication Systems Engineering Dept.
Faculty: Engineering
E-mail: hadar@bgu.ac.il

Extraction of moving objects from a video

In this research we developed an algorithm for extraction of moving objects from the background. The purpose of this work was mainly focused on video compression in the well-known compression standard MPEG-4. Our algorithm is based on a temporal histogram of each pixel value and finding the peak of that histogram. We succeeded in separating the moving objects from the surrounding background in real-time for purposes of efficient video compression. This kind of project is well suited to the homeland security area, especially for detection of suspect objects, or detection and tracking of terrorists.

Compression of Hyperspectral images containing a sub-pixel target

Hyperspectral (HS) image sensors measure the reflectance of each pixel at a large number of narrow spectral bands, resulting in a three-dimensional representation of the captured scene, which consumes a great amount of storage space and transmission time. Hence, the HS images (HSI) representation should be reduced with a compression method which corresponds to the usage and processing of the image. Many compression methods have been offered for HS images, each having its merits and shortcomings. This research focuses on images that contain sub-pixel targets. This target type requires minimum spatial lossy compression, in order to preserve the target, and the maximum possible spectral compression that would still enable target detection.
The proposed compression method for this target type is PCA-DCT (principle component analysis followed by discrete cosine transform). It combines the PCA ability to extract the background in a small number of components, and the individual spectral compression of each pixel of a residual image, using quantized DCT coefficients. The residual image is obtained by subtracting the participating principal components and the mean value of each band from the image. The compression method is kept simple for fast processing and implementation, and considers lossy compression only on the spectral axis.
The goal of this research is to compress a HSI with the minimal possible degradation of point target detection capabilities. It is obvious that different images are affected differently by the same compression depth, depending on several characteristics of the image, first and foremost the amount of details in both spatial and spectral dimensions. Thus, I design a set of spatial and spectral image characteristics that would help determine the appropriate compression depth for the tradeoff between best detection performance and transmission and storage resources. This set of characteristics may also assist in predicting the detection performance of the uncompressed image, since the anomaly detection algorithms are directly related to the background behavior.

Compression of infrared sequences containing a slow moving point target

Infrared imagery sequences are used for detection of moving targets in the presence of evolving cloud clutter or background noise. This research concentrates on slow moving point targets, which are one pixel size, such as long-range aircraft. The infrared (IR) sequences are captured by ground sensors and they contain enormous amounts of data, which transmission to a base unit or storage is very time and resource consuming. Thus, a compression method which maintains the point target detection capabilities is desired. Such a compression was not investigated publicly until now. For this purpose, we introduce a novel parametric compression method, the parabola fit, which utilizes the information about the temporal profile behavior of point targets, clouds and noise pixels, and extracts polynomial parameters from each temporal profile. The second proposed compression method is temporal DCT quantization, which reduces high frequency coefficients. The third and last method is a simple temporal resolution reduction, which discards a constant amount of frames from each group of frames. These three proposed temporal compression methods preserve the temporal profile properties of the point target. These compression methods are evaluated using a SNR-based measure for point target detection. The results indicate that the lossy compression process may improve the detection performance compared to the original movie.

Moving Target Detection

Researcher: Dr. Ofer Levi
Email: Levio@bgu.ac.il

Dr. Levi's main research interests include application of optimization, computational methods and statistics to problems in image processing and analysis in application fields such as moving target detection and compressed sensing imaging systems.