Statistical analysis techniques for aerial infrared images in wavelets transform domain

Abstract

Statistical analysis of airborne IRLS (InfraRed Line-Scanner) images in DWT (Discrete Wavelet Transform) numerical domain is an extension of statistical measurements of computer gener,ated images in our earlier works aimed to finding method for IRLS image coding. The initial results obtained with test images indicate that proposed statistical techniques in the wavelet transform domain are efficient for joint space-frequency 20 anisotropy analysis and adaptive data compression of IRLS images. Image analysis techniques can be classified as phenomenological, structural, deterministic, and statistic. Statistical parameters of image and corresponding radiance and terrain temperature distribution describe basic characteristics of scenes which are of practical importance when modeling scenes, evaluating and predicting IR system performance, developing algorithms for digital processing and data compression for storing or transmission. The statistical properties of high-resolution InfraRed Line-Scanner (IRLS) images of natural terrain in high demand surveillance application are analyzed in this paper. IRLS is a mechanic-optoelectronic device for high-resolution image acquisition in the thermal IR spectral range. The terrain is sampled by a transverse scan as the aircraft moves forward along a track. An obtained series of signals, each corresponding to a terrain scanning line, forms a two dimensional image without temporal information, because each line consists of new data. Consequently, IRLS operates as a linear angle scanning system, giving angular scan and linear displacement along track as the two dimensions of the image. IRLS is more sensitive along the scan direction, giving highly correlated images with the serious geometrical distortion in approaching the horizon [1,2]. Over the past years many classical statistical techniques, in space or frequency domain, which involve probability density function (pdf), autocorrelation funCtion (acf) and power spectral density (psd), have been utilized to analyze linear scan patterns in one dimension (1 D). However, IRLS requires the analysis of two-dimensional (20) aspect of acquired data. Many of the analysis techniques employed for 10 data may be either used directly by averaging over two dimension data or by generalization to a 20 equivalent [3.4]. Three major obstacles in statistical analysis of IRLS images can be identified: nonstationarity, anisotropy and edge sensitivity. An assumption on stationarity of image background, i.e. that the statistical properties do not vary over an image, may be true for a homogenous terrain, but it is rarely valid for real IRLS images, when the image segmentation and isotropy analysis are necessary. The edges in an IRLS image carry very important spatial information about the position and size of objects. In spite of that, there is no good classical measure for the edge sensitivity of an image. One new technique for the analYSis of nonstationary images is DWT (Discrete Wavelet Transform). The DWT offers good edge localization in spatial domain and at the same time good resolution at low frequencies in the transform domain [5], which are desirable properties when analyzing highly correlated IRLS images. Also, the DWT is capable of eliminating the redundant information and providing a compact, multi resolution representation of images for efficient entropy reduction and coding. In this paper, we propose a new application of DWT in the segmentation, edge sensitivity and anisotropy analysis. Various types of IRLS images, carefully segmented along the scan