Multiscale modeling eases target detection
To automatically identify items of interest in imagery or motion video, automatic target recognition (ATR) uses image-analysis algorithms to determine possible target locations and then to test them against specific criteria.
To automatically identify items of interest in imagery or motion video, automatic target recognition (ATR) uses image-analysis algorithms to determine possible target locations and then to test them against specific criteria. In synthetic aperture radar (SAR) imagery, the standard method of target detection compares pixels within an image against a specified threshold. If the intensity exceeds a specified threshold, pixels are defined as targets; if their intensities do not exceed this threshold, they are classed as part of the image background. "Unfortunately," says James Arnold, director of sensor information systems at SRI International Montana Operations (Helena, MT), "this can lead to a large number of false alarms when targets appear in a cluttered, highly correlated background."
SRI multiscale detection algorithm was performed on two nonregistered images of a farm scene (a and b). The results of applying the multiscale detector to an image (a) results in the image having detected objects shown as dark outlines (c). Same scale detection is then performed on an alternate image (b), resulting in the image shown in (d). In the final step, change detection is executed by identifying noncorresponding objects (e).
To overcome this limitation, SRI has developed a technique for finding targets by modeling the image at different spatial resolutions using multiscale statistical modeling (MSM). "Multiscale-statistical-modeling techniques provide an effective and efficient means for size-based target nomination," says Arnold. "Unlike template or wavelet transform-based approaches that attempt to match spatially scaled 2-D basis functions with a target-intensity profile, the MSM approach identifies targets by forming and comparing statistical models of the targets and background clutter."
Multiscale image representations are popular for a variety of reasons. In some cases they can isolate objects or features in an image based on spatial data content. "For example," says Arnold, "the wavelet transform provides a way to split an image into different sub-bands in spatial frequency." Some types of imagery, such as aerial views of natural terrain and foliage, can be modeled as fractals and can be described in terms of their statistical properties as a function of scale. Multiscale representations also often lead to computationally efficient processing algorithms that take advantage of recursion and downsampling in scale.
Researchers at SRI have shown that MSM can successfully identify vehicles in SAR imagery without false alarms caused by natural clutter or building structures. In nonregistered aerial digital images of complex farm scenes, for example, the multiscale detection algorithm is first applied to demarcate regions containing objects larger than 4 x 4 pixels or coarser. Applying the multiscale detector to the image results in detected objects shown as dark outlines. The same scale detection is then performed on the second view of the scene. In the final step, change detection is performed by identifying noncorresponding objects.