articles

This article shows how the modified U-Net, a neural network, can be used to find differences in visible and radio frequency spectrum images. The neural network was modified, with its convolutional layers replaced by the convolution blocks with neuromorphic microbiological cells, which partially destroy the cellular skeletal structure and change their conductivity through controlled biocorrosion. The author developed a method of training a modified neural network based on stimulation of the bacterial layer for the corrosion of conductive components. Functional analysis demonstrated the high efficiency of neural network element configuration and showed that the elements can form interconnected active structures. The author found out that, thanks to the neural network cell’s feature, neutral units can autogenerate signals. This is how information passing through the network can be processed both in passive mode and through interaction with local electrical activity. The author also researched generated activities, which revealed the integral effect of adding signals from neuromorphic cells, resulting in a complex response that includes the spectral components of all neighbouring cells. The modified network has an advantage over similar neural network structures: training can be managed by changing the total activity of neurons, rather than by evaluating the network’s response to test data. When it comes to a trained and formed neural network in which conductive structures are configured, spontaneous activity occurs much less frequently than in the initial configuration where the cells were not subjected to biocorrosion and therefore had maximum conductivity. The experiments demonstrated that the modified U-Net can be used to find differences in visible and radio frequency spectrum images. To successfully find differences hidden by the geometric features of the terrain, the author used a comprehensive strategy for image comparison using visible and radio spectra. The practical research is novel in that it offers a newly developed modification of neuromorphic cells. They achieve high speed of task solution due to the massively parallel organisation of detecting changes in images.