Analysis of Hardware-Implemented U-Net–Like Convolutional Neural Networks

Two convolutional neural networks (CNNs)—modification of classic U-Net and U-Net with the use of dilated convolutions—were implemented. In order to train and test the CNNs, we utilised unmanned aerial vehicle images containing Abies sibirica trees damaged by Polygraphus proximus. The images consist five classes: four classes of the trees depending on their condition and background. The weights of the CNNs, obtained as a result of the training, were then used to implement the CNNs in the field-programmable gate array–based system on a chip platform (Xilinx Zynq 7000). The paper also presents a comparison of the hardware-implemented CNNs in terms of segmentation quality and time efficiency.