Властивості згорткові нейронної мережі на основі автокодера

Abstract

The convolutional properties of the autoencoding neural network for the object detection problem in the image are considered. Data sets in the form of two-dimensional images with three color channels were generated for training and testing. The images are generated based on a three-dimensional scene consisting of objects such as spheres, cubes, cylinders and “monkey” models. The time of network train-ing on the data with different configurations and the result at the output of the neural network were estimated.