The impact of distribution shift on object segmentation accuracy using convolutional neural network
Journal Title: Modern Innovations, Systems and Technologies - Year 2025, Vol 5, Issue 2
Abstract
Neural networks are widely used for object recognition in computer vision. Object recognition accuracy decreases as the statistical parameters of a test dataset deviate from those parameters of a training dataset. Such phenomenon is called a distribution shift. This paper investigates the impact of the distribution shift on object segmentation accuracy in digital images. A remarkable portion of current works in this field simulates the distribution shift by means of special effects applied to real world images. Such approach constraints the range of controllable parameters and reduces the possibilities for distribution shift analysis. We synthesize the initial digital images of geometric primitives in order to estimate the impact of distribution shift on object segmentation accuracy. Our approach allows to gain more control over neural network train and test processes as well as to analyze a wider range of distribution parameters. For current distribution shift analysis we have chosen such image parameters as size of distinct objects, rotation angle of distinct objects and the whole image, image color components, image noise variance, objects mutual occlusion index.
Authors and Affiliations
A. S. Chumakov, I. I. Lychkov
Keywords
Related Articles
- EP ID EP770152
- DOI 10.47813/2782-2818-2025-5-2-3077-3085
- Views 4
- Downloads 0
The problem of the lack of a unified reference application for the labeling system of medicines
A few years ago, an experiment on the labeling of medicines began in Russia. The rapid development of this experiment forced participants in the circulation of medicines to face a number of difficulties. This is caused b...
Brain-computer interfaces: technology overview and modern solutions
The purpose of this study is to provide an overview of the current state of neural interface technology and to compare their various modern implementations with each other, highlighting their advantages and features. The...
Review and comparativeanalysisofBPMN-systemsfor robotization of business processes
The purpose of this article is to give an overview of modern programs in which it is possible to implement the methods of RPA technologies. This article provides a comparative analysis of software within which it is...
Prospects for the development of electronic payment systems
This article examines the prospects for the development of electronic payment systems and their impact on modern society. Key aspects such as security, speed of transactions, accessibility and ease of use are analyzed in...
Static balancing of steel inter-saw gaskets
The results of calculations of static balancing of steel inter-saw gaskets for cotton gins are presented in this article. For manufacturing of steel inter-saw gaskets, a welded connection for the belts is used, which lea...