Background Adaptive Band Selection in a Fixed ... - Semantic Scholar

6 downloads 0 Views 226KB Size Report
Using only a subset of the available bands can decrease false alarms while ... A Constant False Alarm Rate detection algorithm is outlined in Section 2.
Proceedings of the SPIE Conference on Detection and Remediation Technologies for Mines and Minelike Targets VII 4742, April 2002.

Background Adaptive Band Selection in a Fixed Filter System Frank J. Crosby, Harold Suiter Naval Surface Warfare Center, Coastal Systems Station, Panama City, FL 32407 ABSTRACT An automated band selection algorithm suitable for real-time application with fixed filter multispectral cameras is presented for multispectral target detection. Fixed filter multispectral cameras collect all bands regardless of the background. Background adaptive band is the selection of a subset of the bands for target detection processing. Fixed filter systems typically include a small number of general-purpose bands. The bands are chosen to enhance targetbackground contrast but are not keyed to specific target features. In some situations it is unlikely that all bands contribute to target discrimination. Using only a subset of the available bands can decrease false alarms while maintaining target detection performance and reduced processing requirements. The advantages are demonstrated using six band multispectral data and two distinct background categories. Keywords: Background Categorization, Multispectral Target Detection, Fixed Filter Multispectral Camera, False Alarm Reduction, Background Adaptation

1. INTRODUCTION The most effective application of multispectral processing is achieved when the particular spectral bands are chosen to provide the most contrast between targets and clutter. The bands that contribute most to multispectral-based detection depend on both the spectra of the background and that of the target. However, selecting bands based on both targets and backgrounds is a very restrictive condition and the reflectance of targets and backgrounds is highly variable. Restricting bands to account for one target on one background is unduly limited in scope and unfavorable in operational situations. One solution to achieving both improved band selection and general applicability is background adaptive band selection. The dynamic selection of bands in a tunable filter system based only on backgrounds has been shown to significantly improve target detection1. Fixed-filter multispectral cameras are more common and less expensive than tunable multispectral cameras. In a fixed filter system, all bands are collected regardless of the background. In this case, background adaptive band becomes the selection of a subset of the bands for target detection processing. Using only a subset of the available bands can have two advantages. One advantage is a reduction in false alarms while maintaining target detection performance. The contribution of some bands may raise the level of false alarms more than level of correct detections. The other advantage is reduced processing. The degree of reduction is detection-algorithm dependent. A Constant False Alarm Rate detection algorithm is outlined in Section 2. The design of a general-purpose set of filters for a multispectral camera is guided by possible target and background occurrences. The filters are designed to account for as many situations as possible. Therefore, it is likely that some of the bands will not contribute to target detection in certain backgrounds. Section 3 describes the culling of a set of six general-purpose bands for two distinct backgrounds. The application of background-adaptive band selection requires the categorization of background types. The background of each scene must be categorized in order for the proper bands to be selected. This imposes the constraint that a common and limited number of bands are available for both background categorization and target detection. Efficient use of background information requires real-time analysis and implementation. A complete characterization and

Proceedings of the SPIE Conference on Detection and Remediation Technologies for Mines and Minelike Targets VII 4742, April 2002.

classification of the background is not usually practical in a tactical scenario. A background categorization algorithm that is based on background and band characteristics and is additionally achievable in real-time is presented in Section 4. The advantages are demonstrated using six band multispectral data in Section 5. Each of the examples demonstrates an advantage of band selection. The first example is beach scenery that shows improved detection performance. The second is vegetation scenery that shows sustained results with reduced processing requirements.

2. DETECTION ALGORITHM PERFORMANCE The two primary reasons to reduce the number of bands are false alarm reduction and faster processing. False alarms can be introduced by spectral/spatial anomalies that exist in bands that are known to highlight naturally occurring phenomena rather than artificial targets. The contribution to target detection is a function of the discrimination obtained in individual-band processing. The bands in a fixed-filter system are logically chosen to be suitable for a wide variety of backgrounds and targets. In some situations it is improbable that all bands contribute to target discrimination. Although multiband processing with this algorithm adds something that single band does not, the existence of a contribution by a particular band can be ascertained by single band processing. The exploitation of spectral differences is complicated by highly variable target and background signatures. In addition to their spectral characteristics, most objects of interest are man-made and exhibit regular geometrical structures. Unlike spectral signatures, the shape of the target can often be specified with a high degree of confidence. If the resolution is sufficient, one can use the shape to enhance detection or reduce false alarms. The FX target detection algorithm was chosen for this application2. It is an adaptive constant false alarm rate technique that uses spatial and spectral information. The FX algorithm is an adaptive algorithm that owes its success to the limited number assumptions placed on the data. Fewer assumptions allow it to be applicable in many different situations. The FX algorithm is a scalar-valued function of the possibly multi-dimensional quantities: average target intensity, the average background intensity, and the common covariance of the background and target. In p-band multispectral images, T each pixel is a p-dimensional sample that can be represented by the column vector x i = [ x1 x 2 K x p ] . A local

= [x1 Lx N ] . This sample is divided into three disjoint sets, background pixels, B = [ b 0 , K , b N B ] , possible target pixels, T = [ t 0 , K , t N T ] , and guard band pixels that sample of the image of size N is given by the p × N matrix X

likely contain both target and background contributions. The guard band pixels are not used in the detection metric calculation. The mean of the target pixels are given by

∑t

N T = µ T . Let t be a vector of the same length as T each element

NT

of which is identical and equal to

µ T . The mean of the background is given by

N B = µ B . Let b be a vector of

∑b NB

the same length as B each element of which is identical and equal to

µB .

The means of the background and target can be used to compute the common covariance matrix:

S= The detection metric is calculated as

[

]

1 (B − b )T (B − b ) + (T − t )T (T − t ) . N −2

Proceedings of the SPIE Conference on Detection and Remediation Technologies for Mines and Minelike Targets VII 4742, April 2002.

F (X) =

NBNT ( µ B − µ T ) T S −1 ( µ B − µ T ) . N

The detection metric is calculated by convolving the image with masks that isolate the target, background, and guard band sets for each pixel. The relationship among the three masks is shown in Figure 1.

Background Mask

Target Mask

Guard Band

Figure 1: Detection Algorithm Masks

The innermost mask is the target mask. The average target intensity is calculated over this area. The outermost mask is the background mask. It defines the local surroundings of the target. The average background intensity is calculated over this area. The guard band is applied to the area adjacent to the target area. However, the pixel values contained within the guard band are not used in calculating detection values. The guard-band region is used to prevent pixels that are neither completely on the target nor totally contained within the background from biasing the target or background pixel average, thus lowering the overall contrast between the target and the background. Good results have been obtained with a 41 × 41 background, a guard band that is 2 pixels larger than the target mask, and a 3 × 3 target mask. 2.1. False Alarm Reduction The contribution of a band is a factor of the amount of target-background contrast, the variation of the background and the false alarms. The detection algorithm employs the inverse of the covariance estimate, which generally obscures a direct assessment of a band’s contribution. However, if there is no target-background contrast in an added band, then the effect of the additional band is either to add to the confidence of the background covariance estimate or introduce/reinforce false alarms. In order for a band to add to the confidence of the background covariance estimate and not add to the false alarms the band would have to contain no anomalies. Having no anomalies implies that the background is perfectly described by the initial assumptions of the FX algorithm, i.e. a Gaussian distribution. In nature, this rarely occurs. 2.2. Reduced processing Implementing background categorization as a global operator maximizes the computational savings. Although background categorization could be implemented on a local level, reduced processing would not likely be obtained. Local implementation has a higher dependence on preprocessing steps. These preprocessing steps may include fixed noise removal and band registration. The only processing reduction in a local implementation is from the detection algorithm. Figure 2 shows the computational requirements for processing six bands. Total processing includes fixed noise removal, registration, and the FX detection algorithm. Designing background categorization to be invariant to preprocessing maximizes the computational savings.

Proceedings of the SPIE Conference on Detection and Remediation Technologies for Mines and Minelike Targets VII 4742, April 2002.

Figure 2: Computational Requirements

3. BAND SELECTION OVERVIEW Fixed filter systems typically include a small number of general-purpose bands. The bands that are available are primarily selected for target detection. The bands are chosen to enhance target-background contrast but are not keyed to specific target features. As an example, a set of six general-purpose bands for a passive multispectral system covering the visual and near infrared wavelengths may be selected based on three environmental phenomena that are likely to influence littoral reconnaissance. The three phenomena are: the scattering of blue and ultraviolet light in the atmosphere, the blue-green transmission peak of seawater, and the near-infrared (IR) reflectance rise of vegetation. There could be a violet band to take advantage of the diffuse lighting as well as poor camouflaging at this end of the spectrum. Using diffuse lighting allows some penetration into shadowed areas. Two bands centered in the blue-green and yellow-green regions would allow for some penetration of water at various depths. Placing a band before the lower edge of near-IR vegetation rise catches an often present reflectance dip. Selecting the fifth band to be coincident with the near IR rise captures some of the rise of vegetation without duplicating the higher wavelength reflectance, which stays fairly uniform. Finally, one band can be set after the vegetation rise. A target camouflaged against human vision, but not using special night vision defeating paint will look dark compared to vegetation. These six bands can be graphically shown using inverted parabolas to represent bandwidths as in Figure 3. They are selected with the vegetation rise and the maximum transmission of seawater in mind, but they have been mostly chosen for adequate performance of the camera

Proceedings of the SPIE Conference on Detection and Remediation Technologies for Mines and Minelike Targets VII 4742, April 2002.

Figure 3: General-purpose multispectral bands giving both adequate spectral coverage and compensation for camera sensitivity curve

An appropriate subset is determined for several likely background categories. Each subset is selected for its ability to improve detection on a particular background category for the widest variety of target types. The determination of the category subsets depends on the background and the bandcenters and bandwidths. Two divergent backgrounds are beach with little vegetation and a grass field. The reflectance curves from these backgrounds are shown in Figure 4 and Figure 5. These plots provide the mean hyperspectral reflectance curve, the mean+/-standard deviation, and the 90% confidence value of the mean computation. Because of the large number of samples extracted for these statistics, on the order of 600, the mean spectral reflectance is determined with very high confidence. Thus, the two confidence curves in each plot below are very close to the mean spectral reflectance. The sharp changes around 0.7 µm (microns) in graph of the grass reflectance are due to sensor calibration errors. Additionally the sensor is not very accurate above 0.9. The sharp variations depicted between 0.9 microns and 1.0 microns are instrument effects and not real.

Figure 4: Grass reflectance.

Proceedings of the SPIE Conference on Detection and Remediation Technologies for Mines and Minelike Targets VII 4742, April 2002.

Figure 5: Sand reflectance.

Local anomalies on the beach are usually due to sparse areas of vegetation. The vegetation is much darker than the sand in the lower bands. In the near-infrared region of the spectrum, the vegetation begins to blend with the sand. Therefore, the fifth and sixth bands are less susceptible to false alarms than the lower wavelength bands. Although the rise is not fully captured in the fifth band, the vegetation-induced false alarms from a single band will not dominate the detection algorithm when two bands are processed. Large areas of vegetation naturally contain certain plants with different levels of chlorophyll. These factors, along with others such as plant health and season, affect the degree and location of the near IR rise in the reflectance. The interaction between large areas of vegetation and the six bands leads to a high level of variance in the fifth band. The relatively large contrast in the reflectivity of the plants obscures low target-background contrast. As a result, the fifth band does not often contribute to target detection in vegetative environments.

4. BACKGROUND CATEGORIZATION The first stage of applying background-adaptive multispectral band selection is rapid background categorization. The rapid background categorization algorithm has been initially designed for sensors that collect imagery at standard video frame rates. Therefore, the rapid background categorization should be finished computing in less than 150 milliseconds (0.2 second). A background category can be determined for each multispectral image. A flow chart of the process is shown in Figure 6

Figure 6: Background categorization implementation.

Proceedings of the SPIE Conference on Detection and Remediation Technologies for Mines and Minelike Targets VII 4742, April 2002.

Background categories are determined by identifying those areas where reduced band processing is likely to be effective. These areas typically display a high level of uniformity. Complex scenes, such as a shoreline with water, wet sand, dry sand, and vegetation, can be placed into a default category. The default category consists of those scenes that are processed with the full set of available bands. Keeping the categories that utilize reduced band sets simple and distinct reduces negative performance effects. Miscategorizing a scene by placing it into the default category puts the system into its original state. Other types of miscategorization are unlikely with simple and distinct categories. The available bands must be used for both background categorization and target detection. Since target detection is the primary goal, it is assumed that the bands are first chosen to provide satisfactory results in the potential operating environment. The bands selected for use in the background categorization algorithm are those that show the greatest potential for first and second order statistics based categorization. First and second order statistics were chosen due to their relatively fast computation. A decision tree is formed based on the differences and ratios in these statistical measures. Then, the decision tree is developed into a real-time algorithm so that any scene can be rapidly placed into one of the categories. An algorithm for the two elementary categories determined in Section 3 along with the default category is shown in Figure 7. The fundamental principals are the near-IR rise shown by vegetation and the relatively uniform reflectance of sand at those wavelengths. The value of delta for tested data was 0.1. The algorithm, with this value, correctly categorized all of the multispectral images in a set of more than 300 covering a wide range of littoral environments.

Figure 7: Background categorization algorithm.

5. EXPERIMENTAL RESULTS The predicted performance enhancement of the FX detection algorithm on each of the two background categories was verified using available data. Six band multispectral data collected over a typical beach and a grass field were used. The beach scene was an opportunity to demonstrate false alarm reduction. The grass field demonstrated faster processing.

Proceedings of the SPIE Conference on Detection and Remediation Technologies for Mines and Minelike Targets VII 4742, April 2002.

5.1. False Alarm Reduction Six bands of a multispectral image of a beach scene are shown in Figure 8. Target locations are indicated with boxes in the first band. The left side of the images contains a mixture of vegetation and targets. As the band wavelength increases, the vegetation becomes less predominate. In band 6 the vegetation is almost invisible.

Figure 8: Beach scene.

The result of processing bands 5 and 6 compared to the full set of six bands is shown in Figure 9. For the higher detection probabilities, the false alarm rate is clearly reduced.

Figure 9: Beach scene performance.

Proceedings of the SPIE Conference on Detection and Remediation Technologies for Mines and Minelike Targets VII 4742, April 2002.

5.2. Reduced processing Figure 10 shows six bands of a vegetation environment. The targets are highlighted in the first band and have same position in the other bands. In band 5 the targets are nearly invisible.

Figure 10: Grass scene

The result of processing the complement of band 5 compared to the full set of six bands is shown in Figure 11. The performance of the two band sets is almost identical.

Figure 11: Grass scene performance.

Proceedings of the SPIE Conference on Detection and Remediation Technologies for Mines and Minelike Targets VII 4742, April 2002.

6. CONCLUSION The dynamic selection of bands in a tunable filter system based only on backgrounds has been shown to significantly improve target detection. The prevalence and cost of fixed-filter multispectral cameras directed the application of those results to small unchanging set of bands. In a fixed-filter system, all bands are collected regardless of the background. So, background adaptive band selection is reduced to the selection of a subset of the bands for through processing. Using only a subset of the available bands can have two advantages. One advantage is a reduction in false alarms while maintaining target detection performance. The contribution of some bands may raise the level of false alarms more than level of correct detections. The other advantage is reduced processing. The degree of reduction is dependent of the amount of preprocessing and the detection algorithm. Both of these advantages were demonstrated with six band multispectral data. The two backgrounds, sand and vegetation, provided simple and distinct categories. Disadvantages are few. Improper categorization is most likely to result in the use of the full band set, which is the original processing chain.

ACKNOWLEDGEMENTS This work was supported by the Office of Naval Research. The author would like to thank Mr. Ned Witherspoon, the ARORA program sponsor.

REFERENCES 1. Crosby, F.J. et al., “Background Adaptive Multispectral Band Selection”, Proceedings of SPIE Detection and Remediation Technologies for Mines and Minelike Targets VI, 4394 April 2001. 2. Crosby, F. and S. Riley, “Signature Adaptive Mine Detection at a Constant False Alarm Rate”, Proceedings of SPIE Detection and Remediation Technologies for Mines and Minelike Targets IV, April, 2001.

Suggest Documents