{"id":{"repo_id":"rgu","oai_identifier":"oai:rgu-repository.worktribe.com:947825"},"canonical_url":"https://search.dev.ndltd.org/etd/rgu/oai:rgu-repository.worktribe.com:947825","repository":{"repo_id":"rgu","name":"Robert Gordon University","base_url":"https://rgu-repository.worktribe.com/oaiprovider"},"display":{"title":"Learning from small and imbalanced dataset of images using generative adversarial neural networks.","abstract":"The performance of deep learning models is unmatched by any other approach in supervised computer vision tasks such as image classification. However, training these models requires a lot of labeled data, which are not always available. Labelling a massive dataset is largely a manual and very demanding process. Thus, this problem has led to the development of techniques that bypass the need for labelling at scale. Despite this, existing techniques such as transfer learning, data augmentation and semi-supervised learning have not lived up to expectations. Some of these techniques do not account for other classification challenges, such as a class-imbalance problem. Thus, these techniques mostly underperform when compared with fully supervised approaches. In this thesis, we propose new methods to train a deep model on image classification with a limited number of labeled examples. This was achieved by extending state-of-the-art generative adversarial networks with multiple fake classes and network switchers. These new features enabled us to train a classifier using large unlabeled data, while generating class specific samples. The proposed model is label agnostic and is suitable for different classification scenarios, ranging from weakly supervised to fully supervised settings. This was used to address classification challenges with limited labeled data and a class-imbalance problem. Extensive experiments were carried out on different benchmark datasets. Firstly, the proposed approach was used to train a classification model and our findings indicated that the proposed approach achieved better classification accuracies, especially when the number of labeled samples is small. Secondly, the proposed approach was able to generate high-quality samples from class-imbalance datasets. The samples' quality is evident in improved classification performances when generated samples were used in neutralising class-imbalance. The results are thoroughly analyzed and, overall, our method showed superior performances over popular resampling technique and the AC-GAN model. Finally, we successfully applied the proposed approach as a new augmentation technique to two challenging real-world problems: face with attributes and legacy engineering drawings. The results obtained demonstrate that the proposed approach is effective even in extreme cases.","abstract_html":"The performance of deep learning models is unmatched by any other approach in supervised computer vision tasks such as image classification. However, training these models requires a lot of labeled data, which are not always available. Labelling a massive dataset is largely a manual and very demanding process. Thus, this problem has led to the development of techniques that bypass the need for labelling at scale. Despite this, existing techniques such as transfer learning, data augmentation and semi-supervised learning have not lived up to expectations. Some of these techniques do not account for other classification challenges, such as a class-imbalance problem. Thus, these techniques mostly underperform when compared with fully supervised approaches. In this thesis, we propose new methods to train a deep model on image classification with a limited number of labeled examples. This was achieved by extending state-of-the-art generative adversarial networks with multiple fake classes and network switchers. These new features enabled us to train a classifier using large unlabeled data, while generating class specific samples. The proposed model is label agnostic and is suitable for different classification scenarios, ranging from weakly supervised to fully supervised settings. This was used to address classification challenges with limited labeled data and a class-imbalance problem. Extensive experiments were carried out on different benchmark datasets. Firstly, the proposed approach was used to train a classification model and our findings indicated that the proposed approach achieved better classification accuracies, especially when the number of labeled samples is small. Secondly, the proposed approach was able to generate high-quality samples from class-imbalance datasets. The samples' quality is evident in improved classification performances when generated samples were used in neutralising class-imbalance. The results are thoroughly analyzed and, overall, our method showed superior performances over popular resampling technique and the AC-GAN model. Finally, we successfully applied the proposed approach as a new augmentation technique to two challenging real-world problems: face with attributes and legacy engineering drawings. The results obtained demonstrate that the proposed approach is effective even in extreme cases.","abstract_has_math":false,"creators":["Ali-Gombe, Adamu"],"institution":"Robert Gordon University","degree_name":"PhD","degree_level":"Doctoral","degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":["E. Elyan and C. Jayne"],"committee_chairs":[],"committee_members":[],"year":2019,"date_issued":"2019","date_published":"2019","updated_at":"2026-07-24T04:10:00Z","subjects":["Machine learning","Deep learning","Image classification","Class imbalanced data"],"languages":["en"],"rights":[],"rights_urls":[],"identifier_entries":[{"key":"dc:identifier","label":"Identifier","values":["oai:rgu-repository.worktribe.com:947825"],"render_values":[{"text":"oai:rgu-repository.worktribe.com:947825","href":null,"code":true}]},{"key":"dc:creator.authoridentifier","label":"Author Identifier","values":["0000-0002-7925-3347"],"render_values":[{"text":"0000-0002-7925-3347","href":"https://orcid.org/0000-0002-7925-3347","code":true}]}]},"links":{"outbound_url":"https://rgu-repository.worktribe.com/947825/1/ALI-GOMBE%202019%20Learning%20from%20small%20and%20imbalanced","outbound_label":"Repository record","outbound_source":"dc:identifier.uri"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.advisor","label":"Advisor","values":["E. Elyan and C. Jayne"]},{"key":"dc:contributor.sponsor","label":"Sponsor","values":["National Information Technology Agency, Ministry of Communications, Ghana"]},{"key":"dc:creator","label":"Author","values":["Ali-Gombe, Adamu"]},{"key":"dc:creator.authoridentifier","label":"Author Identifier","values":["0000-0002-7925-3347"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2019-12-31"]},{"key":"dc:date.issued","label":"Date","values":["2019"]},{"key":"dc:publisher.institution","label":"Dc Publisher Institution","values":["Robert Gordon University"]},{"key":"dc:relation.isreferencedby","label":"Dc Relation Isreferencedby","values":["https://rgu-repository.worktribe.com/output/947825"]},{"key":"dc:type","label":"Dc Type","values":["Thesis"]},{"key":"dc:type.qualificationlevel","label":"Dc Type Qualificationlevel","values":["Doctoral"]},{"key":"dc:type.qualificationname","label":"Dc Type Qualificationname","values":["PhD"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Machine learning","Deep learning","Image classification","Class imbalanced data"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["en"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["oai:rgu-repository.worktribe.com:947825"]},{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://rgu-repository.worktribe.com/947825/1/ALI-GOMBE%202019%20Learning%20from%20small%20and%20imbalanced"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["The performance of deep learning models is unmatched by any other approach in supervised computer vision tasks such as image classification. However, training these models requires a lot of labeled data, which are not always available. Labelling a massive dataset is largely a manual and very demanding process. Thus, this problem has led to the development of techniques that bypass the need for labelling at scale. Despite this, existing techniques such as transfer learning, data augmentation and semi-supervised learning have not lived up to expectations. Some of these techniques do not account for other classification challenges, such as a class-imbalance problem. Thus, these techniques mostly underperform when compared with fully supervised approaches. In this thesis, we propose new methods to train a deep model on image classification with a limited number of labeled examples. This was achieved by extending state-of-the-art generative adversarial networks with multiple fake classes and network switchers. These new features enabled us to train a classifier using large unlabeled data, while generating class specific samples. The proposed model is label agnostic and is suitable for different classification scenarios, ranging from weakly supervised to fully supervised settings. This was used to address classification challenges with limited labeled data and a class-imbalance problem. Extensive experiments were carried out on different benchmark datasets. Firstly, the proposed approach was used to train a classification model and our findings indicated that the proposed approach achieved better classification accuracies, especially when the number of labeled samples is small. Secondly, the proposed approach was able to generate high-quality samples from class-imbalance datasets. The samples' quality is evident in improved classification performances when generated samples were used in neutralising class-imbalance. The results are thoroughly analyzed and, overall, our method showed superior performances over popular resampling technique and the AC-GAN model. Finally, we successfully applied the proposed approach as a new augmentation technique to two challenging real-world problems: face with attributes and legacy engineering drawings. The results obtained demonstrate that the proposed approach is effective even in extreme cases."]},{"key":"dc:title","label":"Title","values":["Learning from small and imbalanced dataset of images using generative adversarial neural networks."]}]}],"canonical_facts":{"dc:contributor.advisor":["E. Elyan and C. Jayne"],"dc:contributor.sponsor":["National Information Technology Agency, Ministry of Communications, Ghana"],"dc:creator":["Ali-Gombe, Adamu"],"dc:creator.authoridentifier":["0000-0002-7925-3347"],"dc:date":["2019-12-31"],"dc:date.issued":["2019"],"dc:description.abstract":["The performance of deep learning models is unmatched by any other approach in supervised computer vision tasks such as image classification. However, training these models requires a lot of labeled data, which are not always available. Labelling a massive dataset is largely a manual and very demanding process. Thus, this problem has led to the development of techniques that bypass the need for labelling at scale. Despite this, existing techniques such as transfer learning, data augmentation and semi-supervised learning have not lived up to expectations. Some of these techniques do not account for other classification challenges, such as a class-imbalance problem. Thus, these techniques mostly underperform when compared with fully supervised approaches. In this thesis, we propose new methods to train a deep model on image classification with a limited number of labeled examples. This was achieved by extending state-of-the-art generative adversarial networks with multiple fake classes and network switchers. These new features enabled us to train a classifier using large unlabeled data, while generating class specific samples. The proposed model is label agnostic and is suitable for different classification scenarios, ranging from weakly supervised to fully supervised settings. This was used to address classification challenges with limited labeled data and a class-imbalance problem. Extensive experiments were carried out on different benchmark datasets. Firstly, the proposed approach was used to train a classification model and our findings indicated that the proposed approach achieved better classification accuracies, especially when the number of labeled samples is small. Secondly, the proposed approach was able to generate high-quality samples from class-imbalance datasets. The samples' quality is evident in improved classification performances when generated samples were used in neutralising class-imbalance. The results are thoroughly analyzed and, overall, our method showed superior performances over popular resampling technique and the AC-GAN model. Finally, we successfully applied the proposed approach as a new augmentation technique to two challenging real-world problems: face with attributes and legacy engineering drawings. The results obtained demonstrate that the proposed approach is effective even in extreme cases."],"dc:identifier":["oai:rgu-repository.worktribe.com:947825"],"dc:identifier.uri":["https://rgu-repository.worktribe.com/947825/1/ALI-GOMBE%202019%20Learning%20from%20small%20and%20imbalanced"],"dc:language":["en"],"dc:publisher.institution":["Robert Gordon University"],"dc:relation.isreferencedby":["https://rgu-repository.worktribe.com/output/947825"],"dc:subject":["Machine learning","Deep learning","Image classification","Class imbalanced data"],"dc:title":["Learning from small and imbalanced dataset of images using generative adversarial neural networks."],"dc:type":["Thesis"],"dc:type.qualificationlevel":["Doctoral"],"dc:type.qualificationname":["PhD"]},"updated_at":"2026-07-24T04:10:00Z"}