In this test the recognition performance was measured on manually distorted pictures. The pictures in the test set were alter by using a graphic editor. Some of the changed pictures are shown in Figure 7.7.
Figure 7.7: Examples of Manually Distorted Pictures.
The recognition performance of the modular neural network was about 87.8%, the monolithic network could recognize 78.1%. Furthermore the discrimination difference of the modular network was much higher than for the single network. By increasing the demanded confidence (difference between winner and runner-up) the performance difference became even larger.