On the recognition of weakly blurred, highly contrasting objects by neural networks

The influence of the degree of blur on the recognition of high-contrast simple objects is studied using images of handwritten digits and letters as an example. High-contrast objects are objects that are significantly brighter than the surrounding background. White images of handwritten digits (MNIST) and Latin letters (EMNIST) against a black background measuring 28 by 28 pixels are selected as training and testing data sets. It is found that recognition accuracy decreases linearly as the blur level increases. The work also shows that both an increase of blur and a decrease of blur from the blur level of the training sample worsens recognition. It is concluded that recognition when blur is reduced is worse than when blur is increased. Histograms are presented that demonstrate the dependence of recognition accuracy on the degree of blur. It is shown that the initial weights affect the recognition accuracy.