1

page 1 (4 pages)
		2

Proceedings of the Sixth Australian Conference on Neural Networks
Sydney, 6-8 February 1995, pages 152-155.

Baked Product Classification with the Use of a Self-Organising Map

T.RayChaudhuriy J.Chien-Hong Yehy L.G.C.Hameyy C.T. Westcottz

yDepartment of Computing, School of MPCE, Macquarie University, NSW 2109, Australia

zArnott's Biscuits Limited, Homebush, NSW 2140, Australia

Abstract

Study of the baking of biscuits involves among other aspects detailed analysis of colour changes in the product during the process. Previous study has shown the existence of a colour development curve (known as the baking curve) by examining colour development in the RGB and HSI colour spaces.

In the current work a different approach to extracting the baking curve is presented. Using a Kohonen self-organising map with an optimum number of output nodes a well-defined baking curve is automatically extracted from preprocessed data of images gathered during the actual baking process. We propose that these curves can be used as a basis for characterising the colour bake level of a biscuit.

1 Introduction

In order to develop an insight into the process of baking it is important to study the colour changes in the baked product [2]. One approach has been to gather digitized colour image data of biscuits at various stages of baking and to plot the colour pixel values in an RGB or HSI colour cube. Such a plot graphically illustrates how colour development occurs during baking and has been called the baking curve [5].

A method of assessing the bake level with the use of a backpropagation neural network has also been implemented [7]. Intensity histograms of biscuit images have been used as input data to the neural network. The intensity feature of the biscuit images has been used as the criterion for assessment and the backpropagation neural network has been trained to classify these biscuit images. There is scope for improvement in the degree of success of this classification. This is because the classification of biscuits according to bake levels should take into account the overall colour development and not merely the intensity.

The baking curve is a one-dimensional representation of the important colour variations within a threedimensional data space. In order to reduce the dimensionality of data in biscuit colour bake level assessment, measurements should be made along this curve.

We have devised a means to automatically extract the baking curve, using a self-organising map [4]. We suggest that points on the output of such a map should be used as colour bake level bins for preparing a histogram to characterise specific cases of biscuit image data. These histograms would then be used as input to a backpropagation neural network trained to classify them in a manner similar to that in the earlier work on intensity assessment [7].

2 The Self Organising Map

Kohonen's Self Organising Map (SOM) is an unsupervised learning technique that can be used very effectively for extracting structure of complex experimental data. The SOM uses a vector quantisation algorithm that produces a mapping from a high dimensional data space on to a one- or two- dimensional lattice of output nodes. During training the SOM learns the relative ordering of the input data [3]. This kind of network can therefore can be an extremely useful tool for the analysis of complicated experimental data where the data elements bear highly nonlinear relationships to one another.

The SOM networks used in our experiments have three input nodes and between ten and thirty output nodes. The three input nodes represent the threedimensional Red(R), Green(G) and Blue(B) colour components of pixel values from a digitised colour image of a biscuit. The network is designed to have output nodes in the form of a sequential string, i.e., a one dimensional lattice of output nodes and every node in the input layer is connected to each node in the output layer. Such a network will, upon training, map the entire set of RGB pixel values from a biscuit image on to a one dimensional array of points (or output nodes). This has two main effects. Firstly the input data function is compressed into a data space of lower dimensionality and secondly the inter-relationships between the most relevant points in the input data are retained intact in the output format of the network. The network's internal learning rule enables it to extract the inter-relationships without supervision, unlike a backpropagation neural network [1].