Investigation of the influence of the pre-trained bases of neural networks on the quality of segmentation of ore pieces in the photo
Abstract
Investigation of the influence of the pre-trained bases of neural networks on the quality of segmentation of ore pieces in the photo
Incoming article date: 20.10.2023The article deals with the problem of inaccurate allocation of the boundaries of ore pieces after an explosion in a quarry in the photo. In this article, the possibility of using neural networks for segmentation of photographs was investigated, and training, testing and comparison of the pre-trained bases of neural networks were carried out. The family of pre-irradiated bases EfficientNet and SEResNet was tested on the FPN neural network. Neural networks were tested on the same number of learning epochs, and competitively on three, five, seven and ten learning epochs. Training for more than ten epochs was impractical, since almost all networks were undergoing retraining. According to the results of the test and comparison, the result was obtained that the FPN neural network on the pre-trained EfficientNetB2 bases after 7 epochs of training has a segmentation quality of 98.93% in three segmentation classes and 55.1% in the "ore pieces" class.
Keywords: segmentation, neural network, pre-trained foundation, EfficientNet, SEResNet