Visual Segmentation and the Dynamic Binding Problem: Improving the Robustness of an Artificial Neural Network Plankton Classifier

A visual segmentation mechanism for a connectionist pattern recognition system is sought. However, to find such a device requires the solution of the dynamic binding problem. Visual segmentation could be learned by a dynamic binding network. Several puta-tive dynamic binding mechanisms are discussed but each is found to have weaknesses. Two mechanisms are being studied in greater depth so that their weaknesses may be resolved. Also a micro–world for the simulation of visual segmentation tasks is described. 1. INTRODUCTION Simpson et al. (1992) have shown that an artificial neural network (ANN) can discriminate between pre– processed images of Ceratium arcticum (Ehrenberg) and Ceratium longipes (Bailey), two dinoflagellate plankton species. The input patterns on which the network was trained were outline drawings of plankton specimens taken from photomicrographs or camera lucida images. Each outline drawing was digitised and the frequency histogram of the image's power spectrum was determined by a Fast Fourier Transform (FFT). The frequency gradient of the lowest 16 frequency bins of the histogram comprised the input to the network. However, the ANN plankton classifier is not able to classify plankton specimens contained in images where more than one specimen is present (see Figure 1) or where there are also large items of debris such as fragmants of broken plankton or air bubbles. As a result the network is trained and tested with images of single plankton specimens with no large items of clutter. This approach, therefore, fails to address an important challenge for machine vision. This challenge is to enable artificial vision systems to deal with visual images which contain more than one object. In contrast to the ANN plankton classifier, human vision segments an image into its constituent objects. The objective of this study is to improve the robustness of the ANN plankton classifier by incorporating a mechanism which enables recognition of an object separate from the recognition of other objects contained within an image. This mechanism must be able to segment an image into its constituent objects, then generate a representation in which the information relating a single object is grouped together and kept separate from information about other objects. Psychological and connectionist theory will inform the possible mechanisms considered. In particular a general purpose segmentation mechanism is sought, not one that is only able to segment images of plankton. An exciting possibility is a mechanism which learns for itself how to segment images from a …