Human-Centered Computer Vision

Digital media are rapidly replacing their analog counterparts. This development is accompanied by (i) the increasing amount of images present on the Internet, (ii) the availability of the Internet for an increasing number of people, (iii) a decline in digital storage costs, and (iv) the developments in personal digital video/photo cam-era's [1, 2, 3]. In anticipation of these developments, the fields of computer vision (CV) and content-based image retrieval (CBIR) evolved rapidly. Driven by a technology push, a range of CV/CBIR techniques were developed [3, 4]. However, seldomly the user and his characteristics were taken into account and subsequently, limitations of mere technical solutions became apparent [1, 5, 6]. The NWO ToKeN Eidetic project: Intelligent CBIR, aims to bridge the semantic gap present in the field of CV/CBIR, with the successful launch of the CBIR system Vind(X) as its foundation. However, the Vind(X) systems suffers from two drawbacks [7], it depends on: (i) the cooperative annotation of its users to fill its database of outlines [8] and on (ii) outline-outline (or shape) matching [7]. To enable a full analysis of image content (e.g., through object recognition), color and texture analysis has to be done as well as segmentation and shape extraction, to facilitate shape matching. Since each of these topics is essential for CV/CBIR, each of them was addressed in the Eidetic research line and will be discussed, before combining them. Most images present on the Internet and in databases are color images. Moreover, the analysis of color in the image is not only used for the analysis of color distributions but is also used in texture analysis, image seg-mentation, and shape extraction. Hence, color analysis is of the utmost importance, for bridging the semantic gap [5], since color captures essential information about our environment [1, 2, 9]. Therefore, we started with fundamental research toward human color processing [10]. This resulted in a unique color space segmentation, driven by experimental data concerning the 11 color categories, known to be used by humans since half a century [11, 12]. This color space segmentation can function as a highly efficient , human-based color quantization scheme [1, 11]. Texture is the second feature, widely used for image analysis, CV, and CBIR purposes. Most texture analysis techniques are intensity-based [13, 14]. However, multiple problems can arise with texture analysis of color images , when their color is ignored (e.g., two distinct colors can have …