No abstract

We present a so-called Neural Map, a

Visual object understanding [12] relies on the comparison of currently perceived and previously acquired object models, either to discriminate among physically similar ones in the case of recognition, or to generalize common properties across physically different ones during categorization. Current research trends [16] go towards the development of artificial systems that use appearance-based object models combining feature-and correspondence-based approaches to overcome object invariance difficulties. They are structured dynamically using relatively invariant patches of information, which can be shared across objects from the same category. These patches are represented by regular-shaped graphs, whose nodes are labeled with elementary features that capture texture information from the object view. The accumulated world knowledge of these artificial systems is built on a visual dictionary of such patches and must be organized in a way that allows efficient object retrieval and identity or category decisions, respectively. The present work introduces a memory framework for these artificial systems utilizing three different approaches to self-organization of visual object knowledge: the Neural Map, the Neural Map Hierarchy, and the Semantic Correlation Graph. The properties of the proposed computational theory are grounded in the unsupervised structural organization of these object models' components according to their visual resemblance and co-occurrence, as well as in the use of this structure for matching novel components, as illustrated in Figure 1. Both have been identified as important for achieving efficient object recognition and categorization and for providing insight into the knowledge-driven aspect of perception [2]. The Neural Map [3] combines a Growing Neural Gas Network [6] and a classifier inspired by the coding and decoding of information in the brain [15]. A modified winner-take-all voting scheme integrates the feature matching responses of this memory model. The performance for object recognition and categorization of the resulting Neural Map Classifier is validated using image features that derive texture information from object views with different granularity. The present work extends them by employing different representations to encode gray-and color-valued texture information from object views and evaluates alterations on self-organization caused by neural network bootstrapping [1]. The overall results indicate that medium-sized image features with