Introduction
Background
DIRSIG is a synthetic image generator used to model sensor outputs in various image acquisition scenarios. AutoCAD4 wire frame drawings of objects are used in order to construct a scene. These drawings are used to create synthetic images that represent their truth counter parts. Thermal and reflectance properties for various objects are incorporated into the model as well. Finally, using models of the atmosphere (i.e. MODTRAN) and information about the sensor, DIRSIG is able to output a radiance field image that accurately predicts the output radiance found in many natural scenes.
Imaging systems like DIRSIG are beneficial because they allow for a large range of scenarios or setups to be tested with out the need for a large data collection. This creates a reduction in the amount of spending and time necessary to go into the field. DIRISG may also be used to test scenarios that cover a large area (miles) and would be difficult to replicate.
Algorithms such as those used for detection can also be tested on DIRSIG images. DIRSIG images without the sensor in place may be treated than as truth, allowing you to test the abilities of the sensor. Modeling systems like DIRSIG allow you to step backwards and forwards through the imaging system, so that if an error occurs you are aware of what the input was and are able to identify areas that may be malfunctioning. Synthetic image generators also allow for the development and testing of new sensors prior to the initial cost of building them.
The goal of synthetic modeling systems like DIRSIG is to obtain data that is as analogous to true data as possible. In order to obtain realistic results, both spectral and spatial properties of targets must be modeled correctly. The progression of DIRSIG since its first appearance has been to increase the quality of its output. DIRSIG seeks to generate synthetic images that are very similar spatially and spectrally to real objects.
Consider the spatial characteristics of a grass target. The target appears to be textured, but what exactly is it that causes the textured appearance? Individual blades of grass have slightly different reflectance spectra. It is variations in these reflectance spectra that produce the textured appearance of grass. While the spectra are not the same, they are, however, related and contain a certain set of statistics that characterizes a grass target.
In order for DIRSIG to correctly classify a class, it is very important for the model to use the correct statistical relationship for a grass target. Failure to do this may cause noise in the image or result in a misclassification. This research project will examine the current techniques implemented by DIRSIG in characterizing the texture of a grass target.
Texture in DIRSIG is applied to each pixel spectrally3. Currently this is accomplished by examining the image over a certain range of wavelengths to determine which spectra from a database of 100 spectra is a best fit. DIRSIG does this through a mathematical examination of the reflectance spectra in the grass database. The inputted texture image (grayscale) for the grass target is quantified statistically through a measurement of mean and standard deviation in gray value and a z-score is calculated for each of the pixels3.
This information is transformed into reflectance data. Due to their interdependence, z-scores are computed for each grass spectra in the database, for the given range of wavelengths and the closest reflectance spectra is chosen to represent that pixel in the target. This data is then used for any wavelength of the given pixel. This relationship is described in Figure 13.
Figure 13:
Illustration for determining the spectra for a given pixel
While Figure 13 accurately demonstrates the current technique employed by DIRSIG to choose the spectra for a given pixel of the grass target, it also demonstrates the flaw in this technique. Consider the range over which the z-scores are compared. It is in this region that the spectrum with the closest z-score value is chosen. The curve with the closest Z score is now used to describe the entire reflectance spectrum for that pixel, over the entire range of wavelengths.
Figure 13 shows the variations in the different spectra throughout the entire range of wavelengths. It is apparent that while some of the curves have the same shape in the region of interest, they may very immensely in shape in other regions of the spectrum. Using this technique we may choose a reflectance spectrum that looks very much like grass in this bandpass, but does not match the real texture characteristics in the infrared wavelengths.
This is similar to comparing a piece of music at only one frequency. Songs may have the same sound in certain frequencies but sound completely different in many other frequencies. Another description of this is metamerism, looking at a targets color under one type of lighting two colors may look similar, while under different lighting they may look tremendously different.
This project will examine the current techniques that are used by DIRSIG for the texture characterization of grass and determine the quality of these results by comparing them to a truth image of grass. The hypothesis is that the results produced by the current techniques will not be as good as is desired and possible methods for improvement will be examined. The effect of expanding the spectral database for grass targets in DIRSIG will be tested in order to determine if a significant improvement is apparent.
An algorithm will be implemented in DIRSIG that focuses on two additional areas of the reflectance spectrum and compares two additional z-scores when mapping the spectra from the database for each pixel of grass. This method will be tested to determine if it has contributed a significant improvement in the resulting image’s quality. If these two techniques fail to provide significant improvement in quality, further research will be done on texture characterization and implementation for synthetic image generation.