Automatic Design of Locally Adaptive Filters for Pre-processing of Images Subject to Further Interpretation Vladimir V. Lukin 1 , Nikolay N. Ponomarenko 1 , Alexander A. Zelensky 1 Jaakko T. Astola 2 , and Karen O. Egiazarian 2 1 Dept 504, National Aerospace University, Kharkov, Ukraine E-mails: lukin@xai.kharkov.ua , uagames@mail.ru , zelensky@xai.kharkov.ua 2 Institute of Signal Processing, Tampere University of Technology, Tampere, Finland E-mails: jta@cs.tut.fi , karen@cs.tut.fi Abstract Locally adaptive filters are widely used in image processing applications. However, their design commonly requires sufficient efforts and does not take into consideration some important aspects of further processing (interpreting and/or classification) of images. This paper puts forward a novel approach to automatic design of locally adaptive filters subject to further interpretation, namely, detection and localization of small size objects. Design is based on learning with clustering for a test image corrupted by a noise with statistical characteristics observed in real life images to which the obtained filter intend to be further applied. Quantitative data confirming the designed filter efficiency are presented. 1. Introduction Original images obtained by different sensors and imaging systems are commonly corrupted by a noise [1]. Noise type and characteristics can be either a priori known or pre-estimated [2,3]. Often noise presence considerably prevents reliable interpretation and classification of images, in particular, accurate estimation of sensed terrain parameters [4]. Thus, image denoising is a key stage in image processing [4]. A large number of filters have been already proposed (see [5,6] and references therein). Some of them do not require a priori information on noise properties, like a standard median filter [5]. However, a common tendency is that filters, designed for a given type of noise, usually provide better performance. For example, local statistic Lee [7] and Frost [8] filters are intended for processing images corrupted by pure multiplicative noise with a priori known relative variance. Locally adaptive filters (LAFs) have been actively studied recent three decades. In fact, the local statistic Lee [7] and Frost [8] filters are examples of the first locally adaptive filters with the so-called soft switching. The basic motivation for LAF design was that real life images contain considerably different fragments and for their processing in a scanning window fashion it is reasonable to vary filter properties to suit image local behavior in an appropriate manner [6]. In other words, LAFs have demonstrated their advantages in the sense of their ability to provide a trade-off between noise suppression and preservation of edges, details and texture features [9]. However, a LAF design is not a simple task. It often requires a perfect skill of a designer to account for filter, noise and image properties [6]. Early hard switching LAFs (see [6,10] and references therein) had a rather simple structure that included a noise suppressing filter (NSF), a detail preserving filter (DPF) and a local activity indicator (LAI) to be compared with a threshold to perform LAF output switch between NSF and DPF outputs. However, even for such a structure, LAF design requires optimization and/or proper selection of NSF, DPF, LAI, and threshold. Further attempts to improve a performance of LAF were the following: - advanced primary local classifiers instead of LAI with simultaneous increasing the number of different filters to be applied according to the results of primary local classification have been used [11]; - Texture preserving filter (in addition to NSF and DPF) with separation of pixels that belong to texture to one more class was used. This requires a design of special tools for texture detection and localization [10]. Note that reliable primary local recognition or texture detection and localization are very complex tasks. As a result, the obtained classifiers are still not 41 1-4244-0069-4/06/$20.00/©2006 IEEE.