[campus icon] Accesskey [ h ] University of Paderborn - Home
DE deutsch
Die Universität der Informationsgesellschaft

GET-Forschungsseminar Abstracts

Image Segmentation, Feature and Saliency Computation for Artificial Visual Attention based on a Growing Neural Gas

Christian Born, GET Lab

Vortrag: Mi. 17.12.2014, 16:30, Raum P 1.4.17


The concept of saliency is widely recognized to guide the visual attention human observers. Saliency is used in artificial attention systems to pre-select relevant parts of an image for computationally more intensive processing. In this work, a novel artificial visual attention system based on a Growing Neural Gas is proposed. The approach learns graph-based pre-attentional structures and computes color- and shape-based features and subsequently saliencies for them. The proposed method implements concepts traditionally ascribed to saliency. The ability of the proposed system to perform a saliency-based visual selection of conspicuous locations is evaluated in salient object and pop-out detection tasks. The results are compared with those obtained with different artificial attention systems. The proposed system performs well in the salient object detection task. In fact, it is the only assessed method which is able to reach a good performance in both tasks. Additionally, the stability of the estimated features, when the model is applied to dynamic scenes, is assessed.