Color palette generation for image classification by bag-of-colors

There are a large number of colors to represent images (e.g. 256256256 = 16,777,216 colors in an RGB color space) on computers. Since there are too many colors to handle, a large number of colors are reduced by quantization in the image processing in general. When we perform a uniform color quantization, we often get colors which do not fit the real world. Therefore, typical colors should be learned from real world images to generate a practical color palette. The bag-of-visual words based only on local features of grayscale pixel values provides the state of the art technology in the field of image classification, retrieval and recognition. Therefore it is expected to improve the performance by adding the color information to the local features. However, if we increase the number of features to extract from images, it costs memory and time for computation. Moreover, the increase of features affects the performance of recognition. The aim of this paper is to generate appropriate color palette for image classification by the bag-of-colors with less computation time and as few colors as possible to improve the accuracy of image classification.