Subjective Retrieval System for Texture Images Based on Interactions of Orientation and Color

Abstract

We have been researching on modeling of visual cognition and its application to content-based image retrieval. We started with analyzing the relationship between pre-attentive vision and attentive vision. And we focused on human early vision as pre-attentive vision. We selected statistical texture images as targets because attentive vision was hard to act them. Previously, many texture features have been proposed, but most of them were insufficient to account for human subjectivity. And so, we have newly designed texture features which were adequate to account for human subjectivity.

From the viewpoint of physiological fundamentals and psychological review, we have focused on the orientation feature and color feature of an image, and calculating contrast of these features in various resolutions. Next, we have measured psychological responses by using some descriptive adjectives, and corresponded them to our texture features by adopting the canonical correlation statistics. Based on this correspondence, we developed a contrast based subjective texture image retrieval (CBSTIR) system. Our system can retrieve images which give a similar impression, and also predict images by some descriptive adjectives. From the experimental results, we found that approximate color property was significant to subjective retrieval, while the precise color distribution was significant to non-subjective retrieval. Moreover, we found that the global orientation interaction in multi resolutions was significant to subjective retrieval.