Institutional Repository of Vilnius University
Duomenų augmentacijos metodų palyginimas objektų klasifikavimo uždaviniui /
Abstract
dc:descriptionThe main goal of this work is to evaluate and compare the effectiveness of different data augmentation methods for object classification. The tasks of the study included: analysis of the literature to identify relevant and appropriate augmentation methods; application of different augmentation methods to train deep neural networks; evaluation of the classification performance of different deep neural network architectures using different augmentation methods and datasets for training; and a comparison of the performance of the trained classifiers depending on the augmentation methods used. After analysis of literature methods belonging to six different groups – geometric and photometric transformations, erasing and mixing of sample regions, data synthesis with generative adversarial networks and automated augmentation with Faster AutoAugment – were picked for the research. The experiments with different data augmentation methods were conducted using three different deep neural network architectures (ResNet18, MobileNetV2 and BasicNet which was proposed by the author) and three datasets: CIFAR10, Tiny ImageNet and the dataset of malignant and benign skin moles. After conducting the experiments, it was observed that geometric transformations, mixing of samples and Faster AutoAugment methods were effective in improving the performance of classifiers trained with CIFAR10 and Tiny ImageNet datasets. The highest performance boost was observed after applying a mixed augmentation strategy of geometric transformations and CutMix method while training a ResNet18 model, this helped to increase values of accuracy and F1 score by around 20 %. It was also observed that performance of a basic classifier trained with an effective augmentation strategy can exceed that of a more complex architecture trained without applying any augmentation methods. On the other hand, experiments conducted with photometric transformations, region erasing methods and data synthesis using GANs failed to provide evidence that any of these methods are effective in improving the effectiveness of classifiers trained on the given datasets. Experiments conducted with the malignant and benign skin moles dataset produced mixed results, in many cases application of augmentation methods resulted in either precision or recall for the test data being low. This shows that augmentation alone is not always sufficient to improve the performance of the trained classifier.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Medekša, Karolis,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD192842775&prefLang=en_US