Log-map analysis Paolo Gamba a , Luca Lombardi b, * , Marco Porta b a Dipartimento di Elettronica, Via Ferrata 1, 27100 Pavia, Italy b Dipartimento di Informatica e Sistemistica, Via Ferrata 1, 27100 Pavia, Italy article info Article history: Available online 19 September 2008 Keywords: Log-map analysis Pattern matching Vanishing point detection abstract In the literature, different approaches of artificial visual systems inspired to biological solu- tions can be found. In this paper a particular solution will be described with an experimen- tal case study: the foveated retina approach. The solution can be considered as a proposal of ad-hoc hardware implementation of biological inspired solutions, but also as a model for the internal virtual representation of the real world structures. Ó 2008 Elsevier B.V. All rights reserved. 1. Introduction The computer-image analysis process is a long sequence of elementary steps: first of all, images are sequentially stored, pre-processed and segmented, and finally after a feature extraction phase, image content is analysed and interpreted or clas- sified. This open loop paradigm does not support real-time processing, even for the simplest tasks that humans perform in a very easy way. Actually, images are grabbed in a frame memory by commercial cameras at a pixel rate of 10 6 –10 7 pixels per second, implementing a uniform resolution on the camera field of view. Most of the data in memory is useless, and the little produc- tive data has to be extracted with a very heavy computational cost. Conventional computers are unable to manage the prob- lem of the selection of the regions of interest in real time systems. In the last years, several proposals were put towards specialising computer vision systems so that they effectively identify regions and events of interest and manage them in a way similar to biological vision systems. The attention mechanism in humans effectively operates on an even greater amount of data by directing the fovea (the retina part with the greater acuity and the heist resolution) to just the most appropriate regions to accomplish each particular task. The attention is moved in a pre-defined or shortly determined sequence, between the salient areas (scan- path theory). In order to achieve real time performance, in a similar way, computer vision systems must have the following capabilities: they should be able to select ‘on-the-fly’ the regions of interest in space and/or time; they must re-allocate computer resources to maintain the necessary efficiency even if the data load changes considerably following the data set; they have to change the data resolution, as required by the task at hand, in order to operate with local, regional and global features; they must adapt strategies for the analysis of the scene on the basis of the partial results achieved in the previous steps. Among the different proposals that directly address the quoted requirements, some promising proposals are based on the log-map transformations of the spatial domain [1–7]. These approaches consist of analysing the images through a data structure which supplies the scene content at variable resolution. In this way, the exploitation of the most appropriate data-details can be selected for the operation, the task, and the image under analysis. 0167-8191/$ - see front matter Ó 2008 Elsevier B.V. All rights reserved. doi:10.1016/j.parco.2008.09.008 * Corresponding author. E-mail addresses: p.gamba@ele.unipv.it (P. Gamba), luca.lombardi@unipv.it (L. Lombardi), marco.porta@unipv.it (M. Porta). Parallel Computing 34 (2008) 757–764 Contents lists available at ScienceDirect Parallel Computing journal homepage: www.elsevier.com/locate/parco