Method and device which find position probability corresponding to pixel between two sheets of images and appropriate position corresponding to pixel by using it using unsupervised learning of neural network

Abstract

To provide a method and a device which find a position probability corresponding to a pixel with respect to an input image pair by using unsupervised learning of a neural network and a position corresponding to the pixel having favorable resolution and stability by using it.SOLUTION: A method finds a probability indicative of likelihood of a position corresponding to a pixel (hereinafter referred to as a position probability corresponding to the pixel) as output of a node of a neural network. The method finds a sum total of products of a norm of difference of the pixels in the assumed position corresponding to the pixel between input image pairs and the position probability corresponding to the pixel for that case as a pixel deflection reflection term, and finds the sum total of the norms of the difference between filter output and filter input of the position probability corresponding to the pixel as a knowledge reflection term. The method finds the position probability corresponding to the pixel with respect to the input image pair, by using those weight sums as loss when determining a coupling coefficient of the neural network and a threshold value. The method appropriately determines the position corresponding to the pixel between two sheets of input images as a corresponding position in which the probability corresponding to the pixel becomes a maximum in the assumed corresponding position, by using the derived probability corresponding to the pixel.SELECTED DRAWING: Figure 1

Images