Color feature extraction and selection for image retrieval
Chun-Wei Liang, Wen-Yu Chung · 2016
The difficulty in searching natural images is the semantic gap between low-level pixel data and the content that is perceived by human. We observed that when viewing images, people often focus on the major color regions and ignore small, isolated sectors. Thus, instead of considering all pixels in the images, which is usually inefficient and prone to noise, we examine and utilize color features after image segmentation. First, Gaussian smoothing was applied to images which was then followed by region-based segmentation. Clusters which contain high percentage of the total number of pixels were selected and then factored to sixteen predefined color basis vectors. The vectors were then multiplied by the corresponding percentages and the aggregate final color features would represent the image. We used the Corel-1000 database to evaluate our approach. The result indicated that our method has good precision rates in retrieving images in categories of horse, flower and dinosaur.