Toward a perceptual object recognition system

Abstract

[1] demonstrated that humans are easily able to recognize an object in less than 0.5 seconds. Unfortunately, object recognition remains one of the most challenging problems in computer vision. Many algorithms based on local approaches have been proposed in recent decades. Local approaches can be divided in 4 phases: region selection, region appearance description, image representation and classification [2]. Although these systems have demonstrated excellent performance, some weaknesses remain. The first limitation is in the region selection phase. Many existing techniques extract a large number of points/regions of interest. For instance, dense grids contain tens of thousands of points per image while interest point detectors often extract thousands of points. Furthermore, some studies have demonstrated that these techniques were not designed to detect the most pertinent regions for object recognition. There is only a weak correlation between the distribution of extracted points and eye fixations [3]. The second limitation mentioned in the literature concerns the region appearance description phase. The techniques used in this phase typically describe image regions using highdimensional vectors [4]. For example, SIFT, the most popular descriptor for object recognition, produces a 128-dimensional vector per region [5]. The main objective of this thesis is to propose a pipeline for an object recognition algorithm based on human perception which addresses the object recognition system complexity: query run time and memory allocation. In this context, we propose a filter based on a visual attention system [6] to address the problems of extracting a large number of points of interest using existing region selection techniques. We chose to use bottom-up visual attention systems that encode attentional fixations in a topographic map, known as a saliency map. This map serves as basis for generating a mask to select salient points according to human interest, from the points extracted by a region selection technique [7]. Furthermore, we addressed the problem of high dimensionality of descriptors in region appearance phase. We proposed a new hybrid descriptor representing the spatial frequency of some perceptual features, extracted by a visual attention system (color, texture, intensity [8]. This descriptor consist of a concatenation of energy measures computed at the output of a filter bank [9], at each level of the multi-resolution pyramid of perceptual features. This descriptor has the advantage of being lower dimensional than traditional descriptors. The test of our filtering approach, using Perreira da Silva system [10] as a filter on VOC2005, demonstrated that we can maintain approximately the same performance of an object recognition system by selecting only