A Reduced Spherical Model for Optimization of Image Recognition through 3D Color Histograms

Abstract

A 3D color histogram is an image processing technique used to visualize the distribution of colors (Red-Blue-Green) in a picture. Because color distribution does not significantly change if a pictured object is translated or rotated, a 3D color histogram can be used as descriptor for automatic object recognition. However, this task requires cubes with high dimensionality. Within this context, the present work contributes with an approach to reduce the high dimensionality of the 3D color histogram and improve it as descriptor for object recognition. Tests performed with three databases (COIL-100, an own database, and CO3D) and three recognition systems corroborated its suitability for efficient object recognition, achieving overall recognition rates of 97.0% for objects with complex geometry and reflectance features. These results are more competitive when compared with other color descriptors as C-SIFT, RGB-SIFT, Color moments and RGB histograms.