Optimizing 2-Layer Perceptron Neurons Number and Pixel-to-Shift Standard Deviations Ratio for Training on Pixel-Distorted Shifted 60*80 Images in Classifying Shifting-Distorted Objects

Abstract

A problem of classifying shifting-distorted objects is considered. The object model is the flat monochrome 60*80 image of the enlarged English alphabet capital letter. A model of shifting-distorted monochrome images with pixel distortion is developed. The classifier is 2-layer perceptron, whose higher operation speed and poor performance on shifted objects stands against deep neural networks classifying shifted objects well enough with significant delays. The perceptron performance is maximal classification error percentage which depends on two parameters. These ones are the perceptron single hidden layer size and pixel-to-shift standard deviations ratio. The ratio is assumed to advance the perceptron in training on pixel-distorted shifted images in order to classify shifting-distorted objects better. Thus a problem of minimizing a function of two variables is stated, wherein the function is the maximal classification error percentage. Statistically, the optimal hidden layer neurons number is an integer from 390 to 400, and optimal ratio should be set to a value from the segment [0.01; 0.02]. For the accepted object model, these parameters allow to obtain a 2-layer perceptron classifier whose performance is comparable to that one of deep learners. This is about 11.2 % error rate and lower. Compared to results obtained previously, the gain of this two-parameter optimization is at least about 6 %.