MSMW'13, Kharkov, Ukraine, June 23-28 EFFICIENCY OF LOSSY COMPRESSION OF NOISY AND PRE-FILTERED REMOTE SENSING IMAGES V.V. Lukin, A.N. Zemliachenko Dept of Signal Transmission, Reception and Processing, National Aerospace University, 17 Chkalova St, 61070, Kharkov, Ukraine tel. +38 057 7884841, fax +38 057 3151186, e-mails lukin@ai.kharkov.com M.K. Tchobanou Moscow Power Engineering Institute, 111250, Krasnokazarmennaya str. 14, Moscow, Russia E-mail: tchobanou@gmail.com Practically all images acquired by modern remote sensing (RS) sensors as well as other imaging systems are noisy. Noise level and statistics in obtained images depend upon many factors as principle of sensor operation, conditions of imaging, used pre-processing operations (e.g., calibration), etc. Due to this, noise is almost invisible in some images (or sub-bands of multi- or hyperspectral data) whilst it might be quite intensive in other images (or components of multichannel RS data) [1]. One more peculiarity of modern imaging and RS systems is that they often produce a large volume of data [2]. Due to this, there is often a need to carry out compression of RS data. Lossless compression techniques are frequently unable to satisfy requirements to compression ratio (CR). A CR provided by the most efficient lossless coders is only slightly larger than unity especially if noise is quite intensive [2]. Then, one has to deal with lossy compression that has several specific features in the case of dealing with noisy images [2-4]. First, lossy compression applied to noisy images leads to a specific noise filtering effect where optimal operation point (OOP) might exist according to standard and, more rarely, for visual quality criteria [3-5]. Second, lossy compression can be applied to either original (noisy) images or to pre-processed (pre-filtered) images [4]. The latter approach, under certain conditions, has some benefits in terms of conventional quality metrics of compressed images [4, 6]. However, visual quality of images subject to pre-filtering and then to lossy compression has not been analyzed yet. Meanwhile, visual quality of compressed RS data is important for many applications. Thus, the goal of this paper is to compare performance of two procedures (lossy compression of original noisy and pre-filtered images) in terms of visual quality of compressed data using several adequate visual quality metrics. One more peculiarity in lossy compression of noisy images is that quality metrics are calculated in a specific way. Suppose one has a true (noise-free) image { } true ij I , a corresponding noisy image { n ij I }, a filtered image { } f ij I , a lossy compressed noisy image { } nc ij I , and a filtered and then compressed image { }, 1,.., , 1,.., fc ij I i M j N   where M and N define the size of these images. Then, there are several opportunities to calculate any metric that compares a pair of uncompressed and compressed images. The traditional way is to calculate a metric for a compressed image and the corresponding image subject to compression, e.g., { n ij I } and { } nc ij I or { } f ij I and { } f c ij I . However, in fact, we need to have a compressed image as close to { } true ij I as possible. Therefore, it is more reasonable to calculate and to analyze metrics that “compare” compressed images to { } true ij I [3-5]. Certainly, this can be done only for simulated images, i.e. if noise is artificially added to { } true ij I . Then OOP corresponds to such a parameter of compressing method (scaling factor, quantization step, or bpp depending a coder used) for which the difference between the compressed image and { } true ij I is minimal according to a chosen metric. One can argue that any metric for which OOP has to be attained cannot be controlled in practice because { } true ij I is absent. However, our studies have demonstrated [4, 5] that it is possible to provide image lossy compression in the neighborhood of OOP under conditions that a) noise characteristics are known in advance (or pre-estimated with appropriate accuracy) and b) preliminary studies are carried out for test images with producing recommendations on how to set a given coder parameter. In particular, it has been demonstrated [5] that by setting a quantization step (QS) of the coder AGU [7] as 4  standard deviation (σ) of additive noise, one can attain the neighborhood of OOP for practically any image if compression is applied to the noisy image. Noise is assumed zero mean, white and Gaussian. 343 978-1-4799-1068-7/13/$31.00 ©2013 IEEE