Retrieval performance of color descriptors derived from DC components of protected JPEG images

In a content-based image retrieval (CBIR) system, retrieval performance depends on the selected feature. Features extracted from images will be compared during the matching process. In this paper, we evaluate the retrieval performance of color-based descriptors derived from DC images. A DC image is a downscaled image generated by extracting the DC components of a JPEG-compressed image. DC images contain very rich information and can be utilized for various retrieval purposes. In this study, we consider a visually protected image database in which images have been protected by scrambling or encrypting their AC coefficients, while keeping the DC coefficients intact. This framework reduces computational cost and memory usage; therefore, it can be further implemented in a wide range of CBIR systems. We assessed this framework with three common color descriptors: the color layout descriptor (CLD), the dominant color descriptor (DCD), and the dominant color correlogram descriptor (DCCD). The results show that color-based descriptors are suitable for extraction from a DC image. DCD is one of the most powerful descriptors for this framework. In addition, descriptors extracted from a DC image demonstrated different retrieval performance than descriptors extracted from an original image. The effect of image sizes on retrieval performance of color-based descriptors was also investigated. We confirm that image size affects the ranks of retrieved images for some descriptors.