7. Performance Comparison 8. Conclusion

8–2–1 as the third group. The significant difference between the first group and the second group, or the 5–1 topology and the other N–1 topologies, suggests that the grading performance can be significantly reduced when the input dimension is too small. This is because that the important grading criteria is hidden too deeply in the input and is obscured by the other properties [19]. These results indicate that in general, a suitable topology for the colour inspection system must have at least an input dimension of nine, the next higher input dimension tested. Table 2 : P–values from the t–tests on different pairs of the topologies in table 1. The null hypothesis is that the average RMS errors of two topologies on the cross–validation set are equal. The t–tests also indicate that the average performance of ANN with a hidden layer performs significantly better than with no hidden layers. This is shown in the last row of the table 2, where the average RMS error of 8–2–1 topology is significantly different from the other N–1 topologies even at 0.1% significance level. This suggests that the grading of Milk Coffee biscuit is a non-linear problem requiring hidden nodes. Therefore the optimal ANN of all the topologies tested for the grading task is the 8–2–1 topology. Its performance on the test samples is a good indication of the performance of the optimal network on unseen data and this can be compared to the human inspector's performance on unseen data. Our previous work on monochrome images of a different biscuit type tested topologies with up to five hidden nodes and also found that the 8–2–1 is the optimal topology [5]. Also shown in table 1 is the performance of the trained human inspector in classifying the same set of 298 milk–coffee biscuits. The RMS error was calculated from the normalised and averaged scores of the 298 biscuits as described in section 3. The human inspector was well trained for grading Milk Coffee biscuits. The biscuit samples used by the ANN experiments were presented to the inspector during the experiment only. This ensures that the performances of the networks and the human inspector on unseen data are comparable. Since the target values of the test samples used in the ANN experiments were derived from the grades given by the human inspector the ANN's performance on the test sets includes a component of human error. …