Noise immunity of the U-Net fully convolutional neural network model for semantic segmentation of fir trees on UAV images

Abstract: The paper studies the noise immunity of a modified model of a fully convolutional neural network U-Net with robust Cauchy loss function and Focal Loss function when solving the prob-lem of segmentation (pixel classification) of noisy images of fir trees infected by pests. It is shown that the accuracy of mul-ticlassification of such trees decreases with the increase of area noise and the amplitude of impulse noise on the fragments of the training sample. At the same time, the level of accuracy degradation depends on the modified U-Net loss function used for training. For the model with robust Cauchy loss function, there is a slower decrease in noise immunity with increasing values of noise parameters.

Keywords: noise immunity of a modified u-net fully convolutional neural network model, semantic segmentation of fir tree im-ages, unmanned aerial vehicle