Emotional color features for natural image analysis

This paper proposes a new set of image features called emotional color features. They are derived from the relationship between color and emotion. The procedure is as follows. An input color image is transformed from RGB to L*aV and L*C*h color models. Then, the color data are divided into 20 groups of emotional color using the CIE delta E measurement-based color quantization. After that, emotional color features of the input image are calculated from the histogram of the 20 emotional colors. We have applied the proposed features to content-based image retrieval problem in the natural scene image domain. We found that the retrieval results were consistent with users´ color emotional response. The average consistency score of the top ten image retrieval results using the proposed features was at a high level. The proposed features performed better than the traditional color histogram features. In addition, the emotional color features have the desired properties such as size of feature vector is short, and computation is not complex, which are useful for searching huge image databases.