|
The article presents a way of using machine learning algorithms to recognize objects in images. To implement this task, an artificial neural network was used, which has a high adaptability and allows work with a very large set of input data. The neural network was described using a program written in the MATLAB simulation environment. The basic problem faced by the designer of objects recognition is to collect a sufficient training set of images to achieve the high probability of correct recognition. The set of learning patterns in the artificial neural networks may contain from several dozen thousands to one million training samples. In this article at the beginning the neural network was pre-trained trained based on the images included in the publicly available CIFAR 100 database, which are characterized by a small size of 32x32 pixels. It contains 70 000 images assigned to 10 basic categories. Then the author's database, consisting from 1000 pedestrians, cars and road signs was used. The article contains a description of applied algorithm, method of supervised learning and correction of weight coefficients, selection of activation function and operation on max pooling filter. The results of proposed solution are presented in the form of screenshots from calculations and in figures depicting results of recognized objects. Attention was also paid to the impact of used database for learning the network on the speed of calculations and recognition efficiency. The proper selection of number and types of layers, number of neurons, activation function and the value of the learning factor is very important in designing the neural network in application to objects recognition contained in the images. The problems occurring in the process of learning the neural networks and suggestions for their further improvement are also presented.
|
