The perceptron Training Process Optimization for Recognizing Graphic Objects in an Image

The training process of a single-layer perceptron to recognize graphic objects on a pre-segmented image is considered and proposals for the most saving neural network structure are formulated. To describe the images, descriptors invariant to the full group of affine transformations were used, their values were the inputs of a neural network. At the perceptron outputs, the number of graphic objects of each type identified in the image was formed. The training procedure was based on the error back-propagation algorithm. Based on the results of the numerical experiment, the minimum number of the perceptron hidden layer neurons was determined, which guarantees the convergence of the training process, and the training sample volume influence on this amount was revealed. It is shown that for any volume of the training sample, there is a corresponding number of neurons in the perceptron hidden layer, which guarantees a minimum amount of computational work in the process of achieving zero training error.