COMPARISON OF THE PERFORMANCE OF DIFFERENT NEURAL NETWORK ARCHITECTURES AND PRE-TRAINED NEURAL NETWORKS FOR THE CLASSIFICATION OF FOREST FLORA AND FAUNA

Abstract

The aim of this project is to compare the effectiveness of different neural network architectures and pre-trained models for the classification of forest flora and fauna. Experiments were conducted on a dataset containing images of selected plants and animals found in the forest. The project covers the basics of convolutional neural networks, and compares the architectures of the networks used in the study. The results of the experiments compare the effectiveness of different convolutional neural network architectures and pre-trained models, such as EfficientNet or ResNet50. The study includes the training times of selected neural networks, classification times of individual images, and their effectiveness. Additionally, the impact of pre-training networks on the ImageNet dataset on the quality of neural network classification was compared. The purpose of this project was to determine the neural network with the most optimal parameters for use in our robot, �Rumcajs�, which is to be used for monitoring and mapping forest flora and fauna by classifying the segmented images with Segment Anything Model from our advanced camera systems.