Classifying scaled-turned-shifted objects with optimal pixel-to-scale-turn-shift standard deviations ratio in training 2-layer perceptron on scaled-turned-shifted 4800-featured objects under normally distributed feature distortion

Vadim Romanuke

Abstract


A problem of classifying diversely distorted objects is considered. The classifier is 2-layer perceptron capable of classifying greater amounts of objects in a unit of time. That is an advantage of 2-layer perceptron over more complex neural networks like neocognitron, convolutional neural network, deep learning neural networks. Distortion types are scaling, turning, and shifting. The object model is a monochrome 60x80 image of the enlarged English alphabet capital letter. Therefore, there are 26 classes of 4800-featured objects. Training sets have a parameter which is a ratio of pixel-to-scale-turn-shift standard deviations allowing to control normally distributed feature distortion. An optimal ratio is found, at which the 2-layer perceptron performance is still unsatisfactory. Then, the best classifier is further-trained with additional 291 passes of training sets by increasing the training smoothness 10 times more. This favors in decreasing the ultimate classification error percentage from 35.23 % down to 13.91 %. However, the expected practicable distortions are less, so the percentage corresponding to them becomes just 1.85 %, that is only 2 objects of 100 are misclassified. Such solution scheme is directly applied to other classification problems, where number of features is a thousand or a few thousands by a few tens of classes.

Refbacks

  • There are currently no refbacks.


Developed by Institute of Industrial Electronics and Electrical Engineering of RIGA TECHNICAL UNIVERSITY