Development of an Intelligent System for Object Recognition Using Neural Network Ensembles

The study aims to analyze a method for improving the efficiency of image classification systems by combining convolutional neural networks with different architectures into an ensemble of neural networks. The authors have comparatively analyzed the use of neural networks ensembles with different methods of combining them, such as majority vote, averaging, and stacking. The Cifar100 dataset has been chosen for testing. To speed up the development process, the models have not been trained from the very beginning but have been used as feature vector extractors, i.e., one of the types of Transfer Learning, namely, Feature extraction, has been applied. A qualitative and quantitative comparison of the obtained results has been carried out. The final conclusions analyze the possibility of applying the studied ensembles in pattern recognition tasks.