page 1  (4 pages)
2to next section

Copyright ? 1994 J. Yeh and L. Hamey.

Proceedings of the Fifth Australian Conference on Artificial Neural Networks
1994, pages 266-269.

Biscuit Bake Assessment by an Artificial Neural Network

Jeffrey C.H. Yeh, and Leonard G.C. Hamey

Department of Computing

Macquarie University, NSW 2109, Australia


A prototype artificial neural network system for assessing the bake level of biscuits has been implemented. We present performance results and compare the neural network approach with a statistical method and the performance of the trained inspector. The neural network system performs comparably with the other methods.

1 Introduction

Inspection of baked products is very important for food manufacturers, as it ensures correct taste, texture and appearance. This task is normally performed by trained inspectors who examine the product and report unacceptable product. Human inspectors, however, provide subjective judgements that are prone to both shortterm and long-term variations. Digital image processing systems, in contrast, provide objective appearance assessments. When digital image processing is combined with artificial neural networks (ANNs), the resulting system has the potential to learn objective assessment criteria from presentations of acceptable and unacceptable product samples. The experience of a trained inspector can therefore be captured in a machine inspection system, with the benefits of short-term and long-term consistency and reduced operating cost. The trained ANN may be deployed at the immediate post-production inspection point, providing for direct production control. Recently, the food industry has turned to artificial neural networks for product inspection with promising results [4, 5, 7].

We describe a prototype ANN system for the inspection of bake level in biscuits, as indicated by colour development with exposure to heat. In our experiments, one specific product was chosen (figure 1). The product is characterised by regions of high bake colour where a flakey thin blister forms on the top of the biscuit, and regions of low bake colour where blisters do not occur. The positions of the blisters are unpredictable. The inputs to the ANN are preprocessed intensity histograms of the sample images, and the network is trained to as-

Figure 1: Product to be classified.

sess the product bake level based upon classification of the samples by the trained inspector.

2 Sample Preparation

Ninety biscuit samples were collected. Of these samples, thirty were nominally correctly baked, thirty underbaked and thirty overbaked. The samples were digitally imaged after centring them upon a white background (figure 1). Illumination was supplied by two 40W incandescent lamps, one placed to either side of the sample to minimise shadow effects. To eliminate the effect of illumination variation due to the use of AC power, the images were calibrated by a linear transformation of pixel values based on the measured intensity of the white background in each image. An intensity histogram was then produced for each sample image[1]. After imaging, the samples were classified by a trained inspector who separated them into three classes representing underbake, correct bake and overbake. These classification experiments were performed 10