The influence of noise on UK number plate letter recognition in humans and convolutional neural networks
UK number plates contain symbols based on the Charles Wright font. The Driver & Vehicle Licencing Agency (DVLA) regulations specify precise dimensions of the letters, e.g., a uniform width, leading to dissimilar distances between letter features (strokes), presumably resulting in different acuities.
Using the method of constant stimuli with a single interval five alternative forced choice paradigm, we measured acuities for high-contrast DVLA letters (E, S, M, W, O), chosen based on their distinctiveness (O vs. E) and similarity (M vs. W). To assess the influence of noise, we measured acuities for letters with added Salt-and-Pepper noise (10%, 25% and 50%).
Observers showed similar high acuities for the letters E, S, and O but lower acuities for M and W, caused by their high similarity and narrower features. Interestingly, noise levels up to 50% did not significantly impair performance. This invariance to noise motivated a comparison between human performance and that of various deep neural networks used for object classification (e.g., ResNet50). DNNs were trained on 194 alphabets (i.e., 5,044 high-contrast letters), leading to high training accuracy. The trained networks were subsequently tested with DVLA letters, leading to a relatively accurate letter classification for letters without added noise. However, performance dropped significantly for even the smallest noise level (10%). This difference between human and DNN performance in the sensitivity to noise is an interesting example of the different mechanisms underlying letter identification, despite the frequently described similarities between deep convolutional neural networks and the human visual system.
