Using convolutional neural networks in image object recognition and classification tasks
The usage of artificial neural networks in the tasks of object recognition and classification is studied. The classical problem of classifying objects in an image is considered, namely - determining the sex of a person by face. This is due to the fact that there is a successful set of data, which co...
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Zhytomyr Polytechnic State University
2022-06-01
|
| Series: | Технічна інженерія |
| Subjects: | |
| Online Access: | http://ten.ztu.edu.ua/article/view/260832 |
| Summary: | The usage of artificial neural networks in the tasks of object recognition and classification is studied. The classical problem of classifying objects in an image is considered, namely - determining the sex of a person by face. This is due to the fact that there is a successful set of data, which consists of 47,009 images of faces of men and women and 11,649 images of faces for training and testing artificial neural networks. It is suggested to use convolutional neural networks. This approach reduces the amount of information stored in memory, as well as hierarchically separates and aggregates features of the input data. Convolutional neural network consists of several blocks of convolutional and aggregating layers, alignment layer, layers of fully connected neurons, source neuron. The "ReLU" function was selected as the threshold activation function for all neurons except the original one. The activation function of the original neuron is sigmoidal. The neural network was built, trained and tested using the library "TensorFlow", API "Keras.NET", as well as the developed library of methods based on the platform ".NET Standart 2.0». and «WPF». The OxyPlot library was used to build the necessary graphs. The quality of work of the convolutional neural network that depends on the number of blocks and the sizes of the convolution filter has been investigated. The best results are achieved with 3 blocks of convolutional and aggregation layers and a convolutional filter size of 3x3 pixels. The optimum accuracy of image object classification is obtained by training the network for 14 epochs. |
|---|---|
| ISSN: | 2706-5847 2707-9619 |