Dynamic Images Comparison using Siamese Neural Network

Dynamic images represent images that have the same context but different content. This means that they show the same object, but in one of the images, the object is rotated, moved, shrunken, enlarged or there is something in front of the object that partially covers it. It is often necessary to comp...

Full description

Saved in:
Bibliographic Details
Main Authors: Steiner, Domagoj, Vranjes, Mario, Pul, Matija, Herceg, Marijan
Format: Conference Proceeding
Language:English
Subjects:
Artificial neural networks
CGIAD dataset
dynamic image comparison
dynamic images generator
Generators
Heuristic algorithms
Shape
Siamese neural network
Software
Telecommunications
Training
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Dynamic images represent images that have the same context but different content. This means that they show the same object, but in one of the images, the object is rotated, moved, shrunken, enlarged or there is something in front of the object that partially covers it. It is often necessary to compare such dynamic images, for example, when testing the correct operation of dynamic graphic user interfaces (GUI). Since the content of the two dynamic images is not the same, standard computer methods of comparing the two images cannot be applied for reliable automatic testing of the equality of these images. Therefore, for their comparison, it is necessary to apply different methods, such as those based on comparing image features. Since machine learning and neural networks (NNs) have been used more and more recently in tasks related to image processing, as part of this paper, we wanted to examine the possibility of using NNs to compare dynamic images. Therefore, it was necessary to provide adequate datasets of dynamic images of sufficient size for the training of the NN model, for which a dynamic image generator was created within this research, which made this possible. The Siamese neural network (SNN) was then trained using new datasets for the dynamic image comparison tasks. Finally, it was tested on a freely available CGIAD dynamic image dataset on which it achieves a precision of 94.45%, recall of 92.94%, and accuracy of 93.74%. Results show that SNNs could be a good choice for dynamic image comparison tasks, but should be additionally improved to overcome the best state-of-the-art existing solutions based on a feature comparison.
ISSN:1847-358X
DOI:10.23919/SoftCOM58365.2023.10271593