Setting the Hidden Layer Neuron Number in Feedforward Neural Network for an Image Recognition Problem under Gaussian Noise of Distortion

Abstract

There is considered an image recognition problem, defined for the single hidden layer perceptron, fed with 5-by-7 monochrome images on its input under Gaussian noise of their distortion. In this neural network the hidden layer neuron number should be set optimally to maximize its productivity. For minimizing traintime duration and recognition error rate both simultaneously there are suggested two ways of solving the corresponding two-objective minimization problem. One of them deals with equilibrium conception, and the other takes Bernoulli criterion for getting the single minimization problem.

Optimal Training Parameters and Hidden Layer Neuron Number of Two-Layer Perceptron for Generalised Scaled Object Classification Problem

“…The correction is in importing a share of pixel distortion into DSS. This makes training sets milder for 2LP whose performance under normally distributed feature distortions (NDFD) is excellent [18], [19]. PSSDR r defines the share of NDFD.…”

“…Advantages of 2LP are simplicity and high speed of classification [12], [15], [16]. Quick resetting and resource-saving by low CEP are believed to be reachable if 2LP is optimised by its hidden layer neuron number (HLNN) along with the training parameters [12], [17], [18]. The criterion of optimisation is minimisation of CEP, which depends on the training parameters.…”

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

“…Let the usefulness of 2LP (16) during the training be measured with its performance function "mse" according to the sum of squared errors [8], [9], [22], [23]. Finally, having preset the minimum performance gradient to 6 10  , let the number of epochs be 5000 in order to prevent long-dragging convergence of TP and to shorten the ultimate TP period for each pass.…”

“…Before training on STSM6080I NDPD from TS (2), 2LP (6) ( 1 5 ) is initialized with MATLAB function "feedforwardnet" from the Neural Network Toolbox [4], [9], [13] Such a transfer within the Neural Network Toolbox is implemented with logarithmic sigmoid MATLAB transfer function "logsig" , which has the derivative due to its smooth "S-shape" [7], [16], [17]. Then 2LP (15) , (16) having been adapted with weight and bias learning rules by Neural Network Toolbox adapt function "adaptwb", is going to be trained with MATLAB backpropagation training function "traingda" (see its usage in [3], [4], [7], [9], [12], [18], [19]). Function "traingda" will update 4800 300 300 26 300 26 1 448 126…”

“…In article [3], a 2LP was initialized and identified for classifying STSOs by the example of 26 classes' general totality of monochrome 60 × 80 images (M6080Is) by way of training with scaled-turned-shifted monochrome 60 × 80 images (STSM6080Is), whose matrices had been added to the matrices of the values of a normal variate with zero expectation and unit variance (NVZEUV) [4], [7]- [9]. Thus STSM6080Is with normally distributed pixel distortion were formed (STSM6080I NDPD ), feeding the input of the 2LP as an updated TS.…”