*4.6. Summary*

To conclude, Table 9 lists a summary of all total results across all four datasets while comparing our 3-CNN system with the well known Haar Cascade Classifier algorithm.


**Table 9.** Summary of the total results over all four datasets contrasting the Haar and 3-CNN algorithms.

As can be seen, the CNN based system always outperforms the Haar algorithms in all sets, by an amount ranging between 10% to 29% in the F1 metric. This is particularly so in the UBEAR dataset, since the Haar classifier is incapable of modelling the higher variety of internal representations required to properly classify images in that dataset.

Figure 14 shows a summary of these results. It is important to remark that that our proposed system has stable performance figures across the first three datasets, all of which consist of perfect purpose-made ear photography. The results only slightly drop when presented with natural images due to the challenges already described. This is in contrast to the Haar classifier, which has wildly disparate results, demonstrating the large dependency of this system on the particular conditions of one dataset or another.

**Figure 14.** Results of our 3-CNN system compared to the Haar classifier over the various test datasets.
