John A. Saghri

Adaptive two-stage Karhunen-Loeve-transform scheme for spectral decorrelation in hyperspectral bandwidth compression

A computationally efficient adaptive two-stage Karhunen-Loeve transform (KLT) scheme for spectral decorrelation in hyperspectral lossy bandwidth compression is presented. The component decorrelation of the JPEG 2000 (extension 2) is replaced with an adaptive two-stage KLT scheme. The data are partitioned into small subsets. The spectral correlation within each partition is removed via a first-stage KLT. The interpartition spectral correlation is removed using a second-stage KLT applied to the resulting top few sets of equilevel principal component (PC) images. Since only a fraction of each equilevel first-stage PC images are used in the second stage, the KLT transformation matrices will have smaller sizes, leading to further improvement in computational complexity and coding efficiency. The computation of the proposed approach is parametrically quantified. It is shown that reconstructed image quality, as measured via statistical and/or machine-based exploitation measures, is improved by using a smaller partition size in the first-stage KLT. A criterion based on the components of the eigenvectors of the cross-covariance matrix is established to select first-stage PC images, which are used in the second-stage KLT. The proposed scheme also reduces the overhead bits required to transmit the covariance information to the receiver in conjunction with the coding bitstream.

Hausdorff probabilistic feature analysis in SAR image recognition

An automatic target recognition algorithm for synthetic aperture radar (SAR) imagery data is developed. The algorithm classifies an unknown target as one of the known reference targets based on a maximum likelihood estimation procedure. The algorithm helps assess and optimize the favorable effects of multiple image features on recognition accuracy. This study addresses four procedures: (1) feature extraction, (2) training set creation, (3) classification of unknown images, and (4) optimization of recognition accuracy. A three-feature probabilistic method based on extracted edges, corners, and peaks is used to classify the targets. Once the three features are extracted from the target image, binary images are created from each. Training sets, which are used to classify an unknown target, are then created using average Hausdorff distance values for each of the known members of the eight target image types (ZSU-23-4, ZIL131, D7, 2S1, SLICY, BDRM2, BTR60, and T62) included in the publicly available MSTAR test data. The average Hausdorff distance values are acquired from unknown target feature images and are compared to each training set. Each comparison provides the likelihood of the unknown target belonging to one of the eight possible known targets. For each target, eight likelihoods (for eight possible unknown targets) are determined based on the Hausdroff distances and the pre-assigned feature weights. The unknown target is then classified into the target type that has the maximum likelihood estimation value.

KLT/JPEG 2000 multispectral bandwidth compression with region-of-interest prioritization capability

The region of interest (ROI) coding feature of JPEG 2000 image compression standard is extended to multispectral imagery. This is accomplished by enabling ROI capability of JPEG 2000 module in the previously developed Karhunen-Loeve/JPEG 2000 compression of multispectral images. Preliminary results, based on subjective, statistical, and machine-based exploitation measures, show significant improvement in the compression performance. Depending on the ROI/background relative size and the desired quality differential, the improvement in the classification accuracy can increase by as much as one hundred percent without an increase in the bandwidth.

An adaptive two-stage KLT scheme for spectral decorrelation in hyperspectral bandwidth compression

A computationally efficient adaptive 2-stage Karhunen-Loeve Transform (KLT) scheme for spectral decorrelation in hyperspectal lossy bandwidth compression is presented. The component decorrelation of the JPEG 2000 (extension 2) is replaced with the proposed adaptive 2-stage KLT spectral decorrelation scheme. Direct application of a single KLT across the entire set of hyperspectal imagery may not be computationally practical. The proposed scheme alleviates this problem by partitioning the spectral data set into small subsets. The spectral correlation within each partition is removed via the 1st-stage KLT operation. To remove the remaining inter-partition correlation, a 2nd-stage KLT is applied to the top few sets of eaui-level principal component (PC) images from the 1st-stage. The computation savings resulting from 2-stage KLT is parametrically quantified. The proposed adaptive 2-stage KLT uses only a fraction of the equi-level 1st-stage PC images in the 2nd-stage KLT process. This adaptive scheme results in reducing the size of the 2nd-stage KLT transformation matrices and further improvement in computational complexity and coding efficiency. It is shown that reconstructed image quality, as measured via statistical and/or machine-based exploitation measures, is improved by using a smaller partition size in the 1st-stage KLT. A criterion based on the components of the eigenvectors of the cross-covariance matrix is established to identify such 1st-stage PC images. The proposed adaptive spectral decorrelation scheme also reduces the overhead bits required to transmit the covariance matrices, or eigenvectors, along the coding bit stream to the receiver through the downlink channel.

Near lossless transform coding of multispectral images

The authors have extended the previously developed adaptive transform coding based multispectral algorithm for the lossless/virtually lossless case. Preliminary results are encouraging compared with traditional lossless implementation. The primary technical challenge is to compensate for the quantization errors introduced in processing the high dynamic range eigen images.

Three-dimensional target modeling with synthetic aperture radar

Conventional Synthetic Aperture Radar (SAR) offers high-resolution imaging of a target region in the range and cross-range dimensions along the ground plane. Little or no data is available in the range-altitude dimension, however, and target functions and models are limited to two-dimensional images. This paper first investigates some existing methods for the computation of target reflectivity data in the deficient elevation domain, and a new method is then proposed for three-dimensional (3-D) SAR target feature extraction. Simulations are implemented to test the decoupled least-squares technique for high-resolution spectral estimation of target reflectivity, and the accuracy of the technique is assessed. The technique is shown to be sufficiently accurate at resolving targets in the third axis, but is limited in practicality due to restrictive requirements on the input data. An attempt is then made to overcome some of the practical limitations inherent in the current 3-D SAR methods by proposing a new technique based on the direct extraction of 3-D target features from arbitrary SAR image inputs. The radar shadow present in SAR images of MSTAR vehicle targets is extracted and used in conjunction with the radar beam depression angle to compute physical target heights along the range axis. Multiple inputs of elevation data are then merged to forge rough 3-D target models.

A rectangular-fit classifier for synthetic aperture radar automatic target recognition

The utility of a rectangular-fit classifier for Synthetic Aperture Radar Automatic Target Recognition (SAR ATR) is examined. The target is fitted with and modeled as a rectangle that can best approximate its boundary. The rectangular fit procedure involves 1) a preprocessing phase to remove the background clutter and noise, 2) a pose detection phase to establish the alignment of the rectangle via a least squares straight line fitting algorithm, and 3) size determination phase via stretching the width and the height dimensions of the rectangle in order to encapsulate a pre-specified, e.g., 90%, of the points in the target. A training set composed of approximately half the total images in the MSTAR public imagery database are used to obtain and record the statistical variations in the width and height of the resulting rectangles for each potential target. The remaining half of the images is then used to assess the performance of this classifier. Preliminary results using minimum Euclidean and Mahalanobis distance classifiers show overall accuracies of 44% and 42%, respectively. Although the classification accuracy is relatively low, this technique can be successfully used in combination with other classifiers such as peaks, edges, corners, and shadow-based classifiers to enhance their performances. A unique feature of the rectangular fit classifier is that it is rotation invariant in its present form. However, observation of the dataset reveals that in general the shapes of the targets in SAR imagery are not fully rotation invariant. Thus, the classification accuracy is expected to improve considerably using multiple training sets, i.e., one training set generated and used for each possible pose. The tradeoff is the increased computation complexity which tends to be offset by ever increasing efficiency and speed of the processing hardware and software. The rectangular fit classifier can also be used as a pose detection routine and/or in conjunction with other ATR schemes, such as shadow-based ATR, that require an initial pose detection phase prior to matching.

Class-prioritized compression of multispectral imagery data

A joint classification-compression scheme that provides the user with added capability to prioritize classes of interest in the compression process is proposed. The dual compression system includes a primary unit for conventional coding of a multispectral image set followed by an auxiliary unit to code the resulting error induced on pixel vectors that represent classes of interest. This technique effectively allows classes of interest in the scene to be coded at a relatively higher level of precision than nonessential classes. Prioritized classes are selected from a thematic map or directly specified by their unique spectral signatures. Using the specified spectral signatures of the prioritized classes as end members, a modified linear spectral unmixing procedure is applied to the original data as well as to the decoded data. The resulting two sets of concentration maps, which represent classes prioritized before and after compression, are compared and the differences between them are coded via an auxiliary compression unit and transmitted to the receiver along with a conventionally coded image set. At the receiver, the differences found are blended back into the decoded data for enhanced restoration of the prioritized classes. The utility of this approach is that it works with any multispectral compression scheme. This method has been applied to test the imagery from various platforms including NOAA’s AVHRR (1.1 km GSD), and LANDSAT 5 TM (30 m GSD), LANDSAT 5 MSS (79 m GSD).

Early forest fire detection using principal component analysis of infrared video

A land-based early forest fire detection scheme which exploits the infrared (IR) temporal signature of fire plume is described. Unlike common land-based and/or satellite-based techniques which rely on measurement and discrimination of fire plume directly from its infrared and/or visible reflectance imagery, this scheme is based on exploitation of fire plume temporal signature, i.e., temperature fluctuations over the observation period. The method is simple and relatively inexpensive to implement. The false alarm rate is expected to be lower that of the existing methods. Land-based infrared (IR) cameras are installed in a step-stare-mode configuration in potential fire-prone areas. The sequence of IR video frames from each camera is digitally processed to determine if there is a fire within cameras field of view (FOV). The process involves applying a principal component transformation (PCT) to each nonoverlapping sequence of video frames from the camera to produce a corresponding sequence of temporally-uncorrelated principal component (PC) images. Since pixels that form a fire plume exhibit statistically similar temporal variation (i.e., have a unique temporal signature), PCT conveniently renders the footprint/trace of the fire plume in low-order PC images. The PC image which best reveals the trace of the fire plume is then selected and spatially filtered via simple threshold and median filter operations to remove the background clutter, such as traces of moving tree branches due to wind.

Exploitation of target shadows in synthetic aperture radar imagery for automatic target recognition

The utility of target shadows for automatic target recognition (ATR) in synthetic aperture radar (SAR) imagery is investigated. Although target shadow, when available, is not a powerful target discriminating feature, it can effectively increase the overall accuracy of the target classification when it is combined with other target discriminating features such as peaks, edges, and corners. A second and more important utility of target shadow is that it can be used to identify the target pose. Identification of the target pose before the recognition process reduces the number of reference images used for comparison/matching, i.e., the training sets, by at least fifty percent. Since implementation and the computation complexity of the pose detection algorithm is relatively simple, the proposed two-step process, i.e., pose detection followed matching, considerably reduces the complexity of the overall ATR system.