Supervised framework for top-down color interest point detection

Abstract

Selective attention is one of the dominant properties of the biological visual system to locate regions of interest in the scene. This article presents a local feature detector based on selective attention. Most previous approaches are bottom-up and do not consider prior information for known object categories. They detect feature points using spatial information present in the image. In contrast, this article presents an attention inspired algorithm that encapsulates the second-moment matrix-based detector to identify feature points based on some color information selectively. The intuition is to use skin color as a top-down cue to generate interest points, finding a high similarity to known spectra. It has applications in a variety of real-time applications such as image retrieval, gesture classification, virtual reality, etc. The technique is inspired by the human visual perception to gain cognizance of regions based on selective boosting of colors. The properties of color models are used to form a distinctiveness function to suppress unwanted background clutter. A relationship between interest points and salient colors in the image is computed using partial correlations in color derivative space. The system is evaluated on the MSRA dataset commonly used for saliency detection. The experiments are based on finding distinct regions having an affinity towards skin color. The robustness of the algorithm is tested in a realistic scenario by separate training and testing datasets. Experimental results show a high level of repeatability for different noise variations, image compression, and blur. The simplicity, robustness, and efficiency of the technique to locate color interest points make it appropriate for real-time vision systems.