A Statistical Approach to Quantifying Clutter in Hyperspectral Infrared Images

Abstract

A method to quantify clutter in hyperspectral intive basis, frared (HSI) images in a framework similar to work done . the need for a measure to form the basis for a preon single-band images is presented. Hereby, all objects in a processing step to discard images, or make a decision on furscene that may be mistaken for targets by an Automatic Target ther processing, Recognition (ATR) algorithm are considered clutter. A hy-. the same need stated above for a post-processing step to perspectral image contains a number of contiguous discrete determine the reliability of the result of running an ATR on a bands within the spectrum. The aim is to obtain a measure scene, and of complexity for hyperspectral images, which will indicate . the inverse-problem problem of creating clutter on ground the inherent difficulty for an ATR to detect targets. We imscenes. e.g. camouflaging. plemented 129 different image clutter metrics, and computed them for a database of synthesized HSI images. A matched Since it is difficult to capture the multifaceted nature of imfilter ATR was used to determine the amount of clutter in the age clutter in a single number, our aim is to obtain bounds images as a baseline. We developed a method to select a subon the performance of any ATR on a scene based on a genset of the metrics that in combination correlated best with the eral clutter quantification scheme. That is, a high value of amount of clutter in an image, and defined this as the clutter this quantity will indicate that any ATR will produce a high complexity measure (CCM). Multiple runs of this selection false alarm (FA) rate. A low value may, however, not result procedure for different training image groups show a domiin a low FA rate. This will depend on the exact nature of the nance of a further subset ofmetrics that best predict the CCM.ATR. Several attempts have been made to derive such a clutOur results also show that the CCM obtained from a varying ter metric for single band images [1]- [4]. In previous work, number ofrandom sample images generalizes well for the enwe obtained and successfully used a single-band clutter comtire database. plexity measure as a basis for band selection in a multi-band ATR detection scheme in [5].

Display Clutter

“…Other forms of clutter, such as roadside clutter, were excluded. Research on related concepts, such as complexity (Donderi, 2006;Fadiran, Molnar, & Kaplan, 2006), was not included unless the concept was explicitly used in connection with display clutter. The search process initially yielded 269 papers.…”

“…Other algorithms are based primarily on the variability in colors, hues, luminance, and other image features within a given space (e.g., Fadiran et al, 2006;Jansen & van Kreveld, 1998;Kim et al, 2011;Meitzler, Gerhart, & Singh, 1998;Namuduri, Bouyoucef, & Kaplan, 2000;Shirvaikar & Trivedi, 1992). One well-known such method is the Schmieder-Weathersby statistical variance (SV) clutter metric, which consists of dividing an image into grid cells with an area twice that of the target and then taking the root mean squared (RMS) grayscale variance between these different grid cells (Schmieder & Weathersby, 1983).…”

Monotonic correlation analysis of image quality measures for image fusion

“…All of these features have been inspired by work to develop clutter complexity measures metrics that characterize the degree to which the background appears target-like [9]. Ideally, the clutter complexity determines how hard it to detect or classify a target in the scene due to the complexity of the background.…”

“…The next 150 complexity features in Table 2 are derived from single band statistics. Inspired by [9], we apply the single band features separately to the red, green and blue bands and output five statistics for the 3 resulting values: 1) mean, 2) maximum, 3) median, 4) minimum, and 5) maximum-minimum. In addition, we convert the RGB image into a gray scale image [20, page 67] and then evaluate the single band feature.…”

“…We modified 25 single band features to form 150 the features. Most all of these features have been evaluated in [9] for the development of a hyperspectral clutter complexity measure.…”