Cellular Associative Neural Networks for Image Interpretation

IMAGE INTERPRETATION C Orovas, J Austin University of York, UK ABSTRACT This paper describes the architecture and the operation of a neural network based system for image interpretation. The system is based on the use of two models of associative neural networks, ADAM and AURA for image and symbolic processing respectively. Employing characteristics of cellular automata theory and applying ideas from syntactic and structural pattern recognition, it uses a hierarchical approach to learn the structure of images. The hardware implementation of this system is based on the C-NNAP hardware platform. INTRODUCTION The structure of patterns found in images and the relationships among the primitive elements of these patterns are of a great importance in any image understanding system. Structure handling systems are following the syntactic and structural approach for pattern analysis (Fu (1), Bunke and Sunfeliu (2)). Although successfully applied at a number of cases, the sensitivity of parsing and graph matching to noise and errors at input data and the lack of generality and of a robust learning capability are the main drawbacks (Tombre (3)). One could say that the above drawbacks could be easily overcome should the proper neural network architecture be used. The problem which arises in such a case is that of dimensionality (Austin (4)). That is caused because, in typical approaches, the whole of an image is presented to the neural network and, unless a very large number of examples is provided, it is di cult for the network to generalize. As is referred in (4), one solution to this problem is to use a hierarchical approach to learn the structure of images. In that, after the initial labelling, the pattern primitives forming the objects in the image are combined together to form elements of a higher level of hierarchy. Messages are exchanged between these elements and the process of moving to higher levels of hierarchy and message passing is repeated until the nal characterisation of the pattern with object level labels. To be appeared in: 6th Int. Conf. on Image Processing and its Applications, IEE, Dublin, 14-17 July 1997 In order to do this a Cellular Automata (Burks (5)) like architecture is employed and the system consists of a number of units, sites, with each one of them performing a simple task. Each site consists of a number of modules which are using associative neural networks capable for symbolic processing. Each site is called an associative processor and all together form a Cellular Associative Neural Network (CANN). The basic principle of operation is that sites exchange messages about their state with their neighbours. After every iteration each site is aware of the state of more distant neighbours and then, this is re ected in its own state. When every site receives the proper amount of information to enable it reason about the pattern(s) of which the site is part of, then object level label(s) is(are) assigned to the site. To be able to perform these state transitions the system needs to know the appropriate rules. These are created during the learning session where examples of symbolic images and the relevant object level labels are presented to the system. Exploiting the characteristics of the learning algorithm and of the associative memories which are used, the system achieves very fast learning times and the ability to generalize depending on the level of constraint relaxation allowed when recalling. Moreover, the speed of operation can be augmented with the use of the dedicated hardware, C-NNAP (Austin et al (6)).