Trainable Regularization in Dense Image Matching Problems

Abstract

This study examines the development of specialized models designed to solve image-matching problems. The purpose of this research is to develop a technique based on energy tensor aggregation for dense image matching. This task is relevant within the framework of computer systems since image comparison makes it possible to solve current problems such as reconstructing a three-dimensional model of an object, creating a panorama scene, ensuring object recognition, etc. This paper examines in detail the key features of the image matching process based on the use of binocular stereo reconstruction and the features of calculating energies during this process, and establishes the main parts of the proposed method in the form of diagrams and formulas. This research develops a machine learning model that provides solutions to image matching problems for real data using parallel programming tools. A detailed description of the architecture of the convolutional recurrent neural network that underlies this method is given. Appropriate computational experiments were conducted to compare the results obtained with the methods proposed in the scientific literature. The method discussed in this article is characterized by better efficiency, both in terms of the speed of work execution and the number of possible errors. Doi: 10.28991/HIJ-2023-04-03-011 Full Text: PDF