Sustainable, Reliable, 


Image Recognition Accuracy Test 


The test was to establish the accuracy of the image recognition neural network when working on data that the network hadn't seen 



When training Neural networks, it is important to observe the networks characteristics on information that is new to it, since if there was an insufficient number of images that the network was initially trained on, then the network will become overspecialised and identify unique characteristics from the training images as opposed to more general characteristics from the individual characters to be identified


The neural network was given new images that the network had not seen during the training of the network when establishing the weightings on each edge and vertex of the network.  The classifications were then recorded in a results matrix


The test showed a small drop in accuracy when compared to the accuracy on the data it was trained with. This indicates that the network has developed some degree of overspecialisation, however not to a dangerous degree. The results matrix also highlighted some commonly mistaken characters such as S commonly being interpreted as a 5 and Z being interpreted as an N. This indicates that it would be useful for more images to be added into the network training set to minimise the overspecialisation and decrease the probability of an incorrect classification.