Determining the blur factor of handwritten characters using a convolutional neural network

Abstract

The images of handwritten digits and Latin letters from the MNIST and EMNIST datasets are considered. Each image, which has a size of 28x28 pixels, is convolved with a 3x3 matrix. The convolution matrices are symmetric with respect to the central element and are normalized so that all elements are non-negative and their sum is equal to one. Each convolution matrix is ​​characterized by a central element whose value varies from zero to one, indicating the blur factor. The blur matrices are formed randomly according to the uniform distribution of a random variable. Thus, all images of the training and test sets of both datasets have different blur factors. In the next step, a LeNet-5 neural convolutional network is trained to find the blur factor of an image. In cases where the training and test sets are from the same dataset, the accuracy of determining the blur factor is 99.92% for MNIST and 97.95% for EMNIST. The accuracy deteriorates to 90.2% and 85.9% when the training and test sets are from different datasets. The accuracy of predicting the blur factor depending on blur amount is analyzed. It is concluded that the minimum and maximum blur factor values are determined best.