Neural networks for object recognition within compositional hierarchies
Joachim Utans · 1992
The central task of a computer vision system is perception, the recognition of objects depicted in an image, i.e. localizing and identifying interesting entities in the scene. The goal of this thesis is to study the use of optimizing neural networks as applied to problems in computer vision, specifically recognition of objects from structure. This thesis examines the problem of object recognition in a specific domain. The system described is designed to recognize two-dimensional images depicting simple line-drawing like objects. Recognition is defined as localizing and identifying instances of objects in the input image. The work follows Marr in adopting a hierarchical representation of objects as compositions of their parts and thus emphasizes the role of the object's spatial structure as a discriminant for recognition. A compositional hierarchy can be represented as a graph in which nodes correspond to objects or parts and arcs denote which parts belong to which object. The system must build a similar data structure for input data; recognition then can be formulated as a graph matching problem. Recognition must be formulated in a way independent of the actual position of an object in the image. Thus, part coordinates are specified in an object-centered frame of reference and a necessary task of the recognition system is to compute the transformation from image-centered to object-centered coordinates. The graph matching problem is formulated as the minimization of an objective function. A recurrent neural network is used to minimize the objective function which specifies the connection weights of the neural network. While the initial implementation of the system makes use of ad hoc hand-designed objective functions, a later implementation derives the model descriptions from a stochastic forward model of the objects that describes the prototypical appearance of an instance of the object as observed in an image. The system deals with uncertainties regarding the parameters of part primitives as well as structural deformation of object instances. A method for initializing the optimization problem is proposed that relies on the analogy between a scale space hierarchy and the compositional hierarchy used for the data base. The intent is that by ignoring spatial detail at fine scales, one can compute an approximate solution to the problem at the coarsest scale. Then, from this starting point, the final solution can be found more easily.