Object Categorization Using VFA-generated Nodemaps and Hierarchical Temporal Memories
Ádám Csapó, Péter Bárányi, Domonkos Tikk · 2007
Object categorization and recognition have proved to be difficult tasks in artificial intelligence for several decades. With the recent emergence of biologically inspired soft-computing methods, promising results in specialized application domains are more and more common. In this paper, we propose a novel object categorization method based on statistical properties of nodes -derived from the VFA model -and hierarchical temporal memories. A referential categorization method, obtained by feeding grayscale pixel levels to hierarchical temporal memories, is used to evaluate the model's performance. Results show that categorization based on the statistics of nodes seems to yield higher success rates. This is in correspondence with Biederman's conjecture in his theory of recognition by components (RBC), according to which the statistics of nodes, end points and corners carry essential and sufficient information for object recognition [1]. The first section of this paper consists of a brief introduction, in which we restate the formal definition of the VFA model, as well as present its node-filtering applications. This will be followed by a presentation of the HTM theory for size-and orientation-invariant object representation. Finally, we give a detailed case study in which a hierarchical temporal memory is used to distinguish between two, as well as several object categories.