Object Localization Using Input/Output Recursive Neural Networks

Abstract

Localizing objects in images is a difficult task and represents the first step to the solution of the object recognition problem. This paper presents a novel approach to the localization problem based on recursive neural networks (RNNs), In particular, a recursive learning paradigm is proposed to process directed acyclic graphs with labeled edges, and to realize mappings between graphs which are isomorph, i.e. that share the same topology of the links. The RNN model, that assumes a graph-based representation of images, uses a state transition function that depends on the edge labels and is independent from both the number and the order of the children of each node. Moreover, the presence of targets attached to the internal nodes guarantees a fast learning, particularly sensitive to the local features of the graph. Some preliminary experiments, carried out on artificial images created using the COIL collection, are reported, showing very promising results