Advanced biologically plausible algorithms for low-level image processing Valentina I.Gusakova 1 , Lubov N. Podladchikova 1 , Dmitry G. Shaposhnikov 1 , Sergey N.Markin 1 , Alexander V. Golovan 1,2 , Seong Whan Lee 2 1 A.B.Kogan Research Institute for Neurocybernetics, Rostov State University, Russia 2 Center for Artificial Vision Research, Korea University, Korea ABSTRACT At present, in computer vision, the approach based on modeling the biological vision mechanisms is extensively developed. However, up to now, real world image processing has no effective solution in frameworks of both biologically inspired and conventional approaches. Evidently, new algorithms and system architectures based on advanced biological motivation should be developed for solution of computational problems related to this visual task. Basic problems that should be solved for creation of effective artificial visual system to process real world images are a search for new algorithms of low-level image processing that, in a great extent, determine system performance. In the present paper, the results of psychophysical experiments and several advanced biologically motivated algorithms for low-level processing are presented. These algorithms are based on local space-variant filter, context encoding visual information presented in the center of input window, and automatic detection of perceptually important image fragments. The core of latter algorithm (the cascade method) are using local feature conjunctions such as noncolinear oriented segment and composite feature map formation. Developed algorithms were integrated into foveal active vision model, the MARR [18, 22, 23]. It is supposed that proposed algorithms may significantly improve model performance while real world image processing during memorizing, search, and recognition. Keywords: Active vision, low-level processing, facial images, perceptually important regions 1. INTRODUCTION At present, in computer vision, the approach based on modeling the biological vision mechanisms is extensively developed [1- 4. 6, 10, 19]. Earlier [18, 22, 23], the biologically plausible active vision model for Multiresolutional Attentional Representation and Recognition (MARR) was developed. The MARR model is Foveal Vision System imitating space-variant resolution from the fovea to the retinal periphery and foveation at the most informative visual objects. Such systems have evident computational advantages as compared with artificial visual systems based on uniform image representation [25]. The MARR model was examined while processing test facial images and demonstrated high rate of recognition invariantly to scale, rotation, and shift. However, model algorithms should be modified for processing natural facial video images. Face recognition in natural scene is one of important areas in computer vision. It demands a fast and reliable solution of many particular visual tasks, such as face region identification, tracking, recognition, etc. Some of these tasks, in particular, invariant recognition of facial images changing over time have no effective solution in frameworks of both biologically inspired and conventional approaches up to now [9, 12, 14, 15, 17, 21, 24, 27, 30]. Evidently, computational problems of face recognition in natural scene can be solved by means of advanced biologically plausible algorithms and system architectures that may extend advantages of foveal vision. It is widely accepted that the basic problem that should be solved for creation of an effective artificial visual system to process real world facial images is a search for high-performance algorithms for following visual tasks: (1) automatic selection of the most informative image (scene) regions for detailed processing because only these regions are important for robust face recognition [11, 12, 14, 17, 29, 30]; (2) low-level image processing that, in a great extent, determines system performance as a