<title>Image quality measures for performance assessment of compresssion transforms</title>

Abstract

Image compression is increasingly employed in the exploitation of limited communication channel bandwidth and maximization of storage system efficiency. Although traditional measures of signal or image quality (e.g., mean-squared error or MSE) are useful in a communication context, few such measures are evident in the open literature that effectively address implementational issues such as visual quality or suitability for automated target recognition (ATh) of a decompressed image. For example, reconstruction errors such as blocking effect, aliasing or suppression of high frequencies, preservation of image or neighborhood statistics, and quantization-induced perturbation of line segments or arcs can degrade visual image quality. In numerous ATh applications, such error is an important source of ATh algorithm performance degradation. Unfortunately, the majority of published image quality measures (IQMs) tend to be based on ad hoc or first-order models of human visual system (HVS) function. Such models do not necessarily correlate with the appearance of image details (e.g., higherorder models) or mathematical descriptions of ATR filters (e.g., non-HVS models). Additionally, published IQMs tend not to address the primary ATR problem of feature visibility or the secondary problem of estimating image or neighborhood error distributions in decompressed imagery. In this paper, a collection of performance measures and IQMs designed specifically for image compression is presented. Beginning with the customary performance measures of MSE and space-time bandwidth product, we progress to spatial measures such as cutoff frequency, effect of greyscale quantization on variance, measures of texture preservation, and statistics of the modulation transfer function (MTF). Texture is estimated by an area-based variance descriptor derived from the Lorentzian distribution, which has been successfully employed in ATh practice [1]. Measures for disruption of linear features such as line segments and arcs are based on analysis of the Hough transform and spatial linear regression. Performance analysis of each measure is couched in terms of computational cost, sensitivity to error, and relevance to ATh scenarios. Examples are given for template- and codebook-based transforms such as JPEG, VPIC, EPIC, and EBLAST [2-5]. <br/> <br/> Keywords: Image compression, Image quality measures, Error analysis