Cem Direkoglu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Cem Direkoglu has a total of 60 co-authors. Here are the top co-authors by number of shared publications:
- Noel Edward O'COnnor
Dublin City University
11 shared publications - Melike Sah
Computer Engineering Department, Cyprus International University, Nicosia, Trkiye
10 shared publications - Mark S. Nixon
University of Southampton
8 shared publications Iveel Jargalsaikhan
Dublin City University
4 shared publicationsSuzanne Little
Dublin City University
4 shared publicationsH. Demirel
3 shared publications
Latest Publications
- 31 January 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 312
A depthwise convolutional variational autoencoder for anomaly detection in complex traffic scenarios from UAV views
- 31 January 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 312
A depthwise convolutional variational autoencoder for anomaly detection in complex traffic scenarios from UAV views
- Open Access
- 26 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 2784-2799
Optimized KiU-Net: A Convolutional Autoencoder for Retinal Vessel Segmentation in Medical Images
Medical image segmentation plays a central role in enhancing diagnosis, surgical planning, and treatment strategies, and in this work, the focus is on segmenting retinal vessels from color fundus images. Precise vessel extraction is essential because vessel morphology reflects several ophthalmic conditions. Wide neural networks such as U Net and KiU-Net have improved retinal vessel segmentation; however, thin and low-contrast vessel regions remain difficult to capture. U Net follows an undercomplete design that limits its ability to retain fine structures, whereas KiU Net combines undercomplete and overcomplete paths to provide better detail extraction but still suffers from accuracy limitations and increased computational cost. We present an Optimized KiU Net model that improves the segmentation of thin and low-contrast retinal vessels while keeping the model lightweight. The design refines convolution channel selection and increases encoder depth, and the final feature fusion uses a single concatenation step, which supports faster convergence with fewer parameters. On the RITE dataset, the model achieves an F1 score of 79.80 and an IoU of 66.30, outperforming U Net, KiU-Net, and other similar sized architectures. Compared to KiU Net, the gains are about four points in F1 and six points in IoU with fewer parameters. Additional evaluation on the GlaS dataset shows F1 and IoU scores of 82.21 and 71.03, demonstrating that the method remains effective when compared with existing approaches of comparable scale.
- 26 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 2784-2799
Open AccessOptimized KiU-Net: A Convolutional Autoencoder for Retinal Vessel Segmentation in Medical Images
Medical image segmentation plays a central role in enhancing diagnosis, surgical planning, and treatment strategies, and in this work, the focus is on segmenting retinal vessels from color fundus images. Precise vessel extraction is essential because vessel morphology reflects several ophthalmic conditions. Wide neural networks such as U Net and KiU-Net have improved retinal vessel segmentation; however, thin and low-contrast vessel regions remain difficult to capture. U Net follows an undercomplete design that limits its ability to retain fine structures, whereas KiU Net combines undercomplete and overcomplete paths to provide better detail extraction but still suffers from accuracy limitations and increased computational cost. We present an Optimized KiU Net model that improves the segmentation of thin and low-contrast retinal vessels while keeping the model lightweight. The design refines convolution channel selection and increases encoder depth, and the final feature fusion uses a single concatenation step, which supports faster convergence with fewer parameters. On the RITE dataset, the model achieves an F1 score of 79.80 and an IoU of 66.30, outperforming U Net, KiU-Net, and other similar sized architectures. Compared to KiU Net, the gains are about four points in F1 and six points in IoU with fewer parameters. Additional evaluation on the GlaS dataset shows F1 and IoU scores of 82.21 and 71.03, demonstrating that the method remains effective when compared with existing approaches of comparable scale.
- Open Access
- 23 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 132194-132208
Quantum Vision Theory in Deep Learning for Object Recognition
We introduce a new perspective and a theory, called quantum vision (QV) theory in deep learning, for object recognition. The proposed theory is based on particle-wave duality of quantum physics. In quantum-scale, an object appears as a wave until it is observed, but after observation the object collapses into a solid object, called particle. In quantum world, every object has a ’wave function’ that contains all the information about the object permitted by the uncertainty principle. Quantum-scale world looks different from our human-scale world. Attempts to relate the microscopic quantum world to our macroscopic world led to philosophical issues and questions. But what if the objects in human-scale world such as cats, dogs and bicycles have wave functions as well?And what will happen if we feed waves of objects to Deep Neural Networks (DNN) instead of collapsed still images of objects captured by cameras? This is the main contribution our work. Inspired from quantum physics, we introduce a new perspective and theory, called Quantum Vision (QV) theory in deep learning that is a completely new perspective for object recognition. The proposed QV theory takes captured still images of objects, and converts them to information wave functions using a deep learning block that is called QV block. The proposed QV block is integrated into sequential CNNs, vision transformers and convolutional vision transformer to generate QV model variants for object classification. Extensive experiments are carried out on several datasets, and results demonstrate that QV model variants perform consistently better than standalone versions.
- 23 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 132194-132208
Open AccessQuantum Vision Theory in Deep Learning for Object Recognition
We introduce a new perspective and a theory, called quantum vision (QV) theory in deep learning, for object recognition. The proposed theory is based on particle-wave duality of quantum physics. In quantum-scale, an object appears as a wave until it is observed, but after observation the object collapses into a solid object, called particle. In quantum world, every object has a ’wave function’ that contains all the information about the object permitted by the uncertainty principle. Quantum-scale world looks different from our human-scale world. Attempts to relate the microscopic quantum world to our macroscopic world led to philosophical issues and questions. But what if the objects in human-scale world such as cats, dogs and bicycles have wave functions as well?And what will happen if we feed waves of objects to Deep Neural Networks (DNN) instead of collapsed still images of objects captured by cameras? This is the main contribution our work. Inspired from quantum physics, we introduce a new perspective and theory, called Quantum Vision (QV) theory in deep learning that is a completely new perspective for object recognition. The proposed QV theory takes captured still images of objects, and converts them to information wave functions using a deep learning block that is called QV block. The proposed QV block is integrated into sequential CNNs, vision transformers and convolutional vision transformer to generate QV model variants for object classification. Extensive experiments are carried out on several datasets, and results demonstrate that QV model variants perform consistently better than standalone versions.
- 23 December 2024
- Published by Springer Nature in Arabian Journal for Science and Engineering
- Vol. 50 (21) , 17299-17316
Precise State-of-Charge Estimation in Electric Vehicle Lithium-Ion Batteries Using a Deep Neural Network
- 23 December 2024
- Published by Springer Nature in Arabian Journal for Science and Engineering
- Vol. 50 (21) , 17299-17316
Precise State-of-Charge Estimation in Electric Vehicle Lithium-Ion Batteries Using a Deep Neural Network
- Open Access
- 11 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 149221-149237
Hybrid Transformer Architectures With Diverse Audio Features for Deepfake Speech Classification
The rise of synthetic speech technologies has triggered growing concerns about the increasing difficulty in distinguishing between real and fake voices. In this context, we propose novel hybrid transformer-based models together with different audio feature analysis techniques and achieved the state-of-the-art results. To the best of our knowledge, none of the existing methods have considered combining various hybrid transformer models together with different audio features for fake speech classification, which forms the main novelty of our work. In our work, transformer models are compared with hybrid transformer architectures including Convolutional Neural Network (CNN)-Transformer (i.e., ResNet34- Transformer and VGG16-Transformer models), Bi-directional Long Short-Term Memory (Bi-LSTM)- Transformer, and Transformer with Support Vector Machine (SVM) using different audio feature extraction techniques. In our approach, we utilize three audio attribute extraction techniques: Mel spectrogram (Mel), Mel Frequency Cepstral Coefficient (MFCC), and Short-Time Fourier Transform (STFT) as input representations. The results of our evaluation with instances of real and fake speech using the ASVspoof LA dataset with hybrid transformer models across various audio features indicate that the STFT feature performs best with the ResNet34-Transformer model, achieving a state-of-the-art performance with a development set equal error rate (EER) of 0.0% and evaluation set EER of 3.22%, surpassing all other methods. In terms of accuracy, the STFT feature also performed best with the VGG16-Transformer model, achieving a development set accuracy of 99.55% and an evaluation set accuracy of 94.04%. These results indicate that the proposed study achieved better performance compared to the baseline and state-of-the-art methods.
- 11 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 149221-149237
Open AccessHybrid Transformer Architectures With Diverse Audio Features for Deepfake Speech Classification
The rise of synthetic speech technologies has triggered growing concerns about the increasing difficulty in distinguishing between real and fake voices. In this context, we propose novel hybrid transformer-based models together with different audio feature analysis techniques and achieved the state-of-the-art results. To the best of our knowledge, none of the existing methods have considered combining various hybrid transformer models together with different audio features for fake speech classification, which forms the main novelty of our work. In our work, transformer models are compared with hybrid transformer architectures including Convolutional Neural Network (CNN)-Transformer (i.e., ResNet34- Transformer and VGG16-Transformer models), Bi-directional Long Short-Term Memory (Bi-LSTM)- Transformer, and Transformer with Support Vector Machine (SVM) using different audio feature extraction techniques. In our approach, we utilize three audio attribute extraction techniques: Mel spectrogram (Mel), Mel Frequency Cepstral Coefficient (MFCC), and Short-Time Fourier Transform (STFT) as input representations. The results of our evaluation with instances of real and fake speech using the ASVspoof LA dataset with hybrid transformer models across various audio features indicate that the STFT feature performs best with the ResNet34-Transformer model, achieving a state-of-the-art performance with a development set equal error rate (EER) of 0.0% and evaluation set EER of 3.22%, surpassing all other methods. In terms of accuracy, the STFT feature also performed best with the VGG16-Transformer model, achieving a development set accuracy of 99.55% and an evaluation set accuracy of 94.04%. These results indicate that the proposed study achieved better performance compared to the baseline and state-of-the-art methods.
- Open Access
- 1 August 2024
- Published by SPIE-Intl Soc Optical Eng in Journal of Biomedical Optics
- Vol. 29 (08) , 080502-8
Convolutional neural network-based regression analysis to predict subnuclear chromatin organization from two-dimensional optical scattering signals
SignificanceAzimuth-resolved optical scattering signals obtained from cell nuclei are sensitive to changes in their internal refractive index profile. These two-dimensional signals can therefore offer significant insights into chromatin organization.AimWe aim to determine whether two-dimensional scattering signals can be used in an inverse scheme to extract the spatial correlation length ℓc and extent δn of subnuclear refractive index fluctuations to provide quantitative information on chromatin distribution.ApproachSince an analytical formulation that links azimuth-resolved signals to ℓc and δn is not feasible, we set out to assess the potential of machine learning to predict these parameters via a data-driven approach. We carry out a convolutional neural network (CNN)-based regression analysis on 198 numerically computed signals for nuclear models constructed with ℓc varying in steps of 0.1 μm between 0.4 and 1.0 μm, and δn varying in steps of 0.005 between 0.005 and 0.035. We quantify the performance of our analysis using a five-fold cross-validation technique.ResultsThe results show agreement between the true and predicted values for both ℓc and δn, with mean absolute percent errors of 8.5% and 13.5%, respectively. These errors are smaller than the minimum percent increment between successive values for respective parameters characterizing the constructed models and thus signify an extremely good prediction performance over the range of interest.ConclusionsOur results reveal that CNN-based regression can be a powerful approach for exploiting the information content of two-dimensional optical scattering signals and hence monitoring chromatin organization in a quantitative manner.
- 1 August 2024
- Published by SPIE-Intl Soc Optical Eng in Journal of Biomedical Optics
- Vol. 29 (08) , 080502-8
Open AccessConvolutional neural network-based regression analysis to predict subnuclear chromatin organization from two-dimensional optical scattering signals
SignificanceAzimuth-resolved optical scattering signals obtained from cell nuclei are sensitive to changes in their internal refractive index profile. These two-dimensional signals can therefore offer significant insights into chromatin organization.AimWe aim to determine whether two-dimensional scattering signals can be used in an inverse scheme to extract the spatial correlation length ℓc and extent δn of subnuclear refractive index fluctuations to provide quantitative information on chromatin distribution.ApproachSince an analytical formulation that links azimuth-resolved signals to ℓc and δn is not feasible, we set out to assess the potential of machine learning to predict these parameters via a data-driven approach. We carry out a convolutional neural network (CNN)-based regression analysis on 198 numerically computed signals for nuclear models constructed with ℓc varying in steps of 0.1 μm between 0.4 and 1.0 μm, and δn varying in steps of 0.005 between 0.005 and 0.035. We quantify the performance of our analysis using a five-fold cross-validation technique.ResultsThe results show agreement between the true and predicted values for both ℓc and δn, with mean absolute percent errors of 8.5% and 13.5%, respectively. These errors are smaller than the minimum percent increment between successive values for respective parameters characterizing the constructed models and thus signify an extremely good prediction performance over the range of interest.ConclusionsOur results reveal that CNN-based regression can be a powerful approach for exploiting the information content of two-dimensional optical scattering signals and hence monitoring chromatin organization in a quantitative manner.
- 22 February 2024
- Published by Springer Nature
- p. 262-270
Multi-class Classification of Voice Disorders Using Deep Transfer Learning
- 22 February 2024
- Published by Springer Nature
- p. 262-270
Multi-class Classification of Voice Disorders Using Deep Transfer Learning
- 29 December 2023
- Published by Springer Nature in Arabian Journal for Science and Engineering
- Vol. 49 (9) , 12091-12111
Identification of Locations in Mecca using Image Pre-Processing, Neural Networks and Deep Learning
- 29 December 2023
- Published by Springer Nature in Arabian Journal for Science and Engineering
- Vol. 49 (9) , 12091-12111
Identification of Locations in Mecca using Image Pre-Processing, Neural Networks and Deep Learning
- 23 December 2023
- Published by Springer Nature
- p. 373-383
Optimized KiU-Net: Lightweight Convolutional Neural Network for Retinal Vessel Segmentation in Medical Images
- 23 December 2023
- Published by Springer Nature
- p. 373-383
Optimized KiU-Net: Lightweight Convolutional Neural Network for Retinal Vessel Segmentation in Medical Images
- Top Cited PapersOpen Access
- 22 September 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 106620-106649
A Survey of Audio Classification Using Deep Learning
Deep learning can be used for audio signal classification in a variety of ways. It can be used to detect and classify various types of audio signals such as speech, music, and environmental sounds. Deep learning models are able to learn complex patterns of audio signals and can be trained on large datasets to achieve high accuracy. To employ deep learning for audio signal classification, the audio signal must first be represented in a suitable form. This can be done using signal representation techniques such as using spectrograms, Mel-frequency Cepstral coefficients, linear predictive coding, and wavelet decomposition. Once the audio signal is represented in a suitable form, it can then be fed into a deep learning model. Various deep learning models can be utilized for audio classification. We provide an extensive survey of current deep learning models used for a variety of audio classification tasks. In particular, we focus on works published under five different deep neural network architectures, namely Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Autoencoders, Transformers and Hybrid Models (hybrid deep learning models and hybrid deep learning models with traditional classifiers). CNNs can be used to classify audio signals into different categories such as speech, music, and environmental sounds. They can also be used for speech recognition, speaker identification, and emotion recognition. RNNs are widely used for audio classification and audio segmentation. RNN models can capture temporal patterns of audio signals and be used to classify audio segments into different categories. Another approach is to use autoencoders for learning the features of audio signals and then classifying the signals into different categories. Transformers are also well-suited for audio classification. In particular, temporal and frequency features can be extracted to identify the characteristics of the audio signals. Finally, hybrid models for audio classification either combine various deep learning architectures (i.e. CNN-RNN) or combine deep learning models with traditional machine learning techniques (i.e. CNN-Support Vector Machine). These hybrid models take advantage of the strengths of different architectures while avoiding their weaknesses. Existing literature under different categories of deep learning are summarized and compared in detail.
- 22 September 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 106620-106649
Top Cited PapersOpen AccessA Survey of Audio Classification Using Deep Learning
Deep learning can be used for audio signal classification in a variety of ways. It can be used to detect and classify various types of audio signals such as speech, music, and environmental sounds. Deep learning models are able to learn complex patterns of audio signals and can be trained on large datasets to achieve high accuracy. To employ deep learning for audio signal classification, the audio signal must first be represented in a suitable form. This can be done using signal representation techniques such as using spectrograms, Mel-frequency Cepstral coefficients, linear predictive coding, and wavelet decomposition. Once the audio signal is represented in a suitable form, it can then be fed into a deep learning model. Various deep learning models can be utilized for audio classification. We provide an extensive survey of current deep learning models used for a variety of audio classification tasks. In particular, we focus on works published under five different deep neural network architectures, namely Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Autoencoders, Transformers and Hybrid Models (hybrid deep learning models and hybrid deep learning models with traditional classifiers). CNNs can be used to classify audio signals into different categories such as speech, music, and environmental sounds. They can also be used for speech recognition, speaker identification, and emotion recognition. RNNs are widely used for audio classification and audio segmentation. RNN models can capture temporal patterns of audio signals and be used to classify audio segments into different categories. Another approach is to use autoencoders for learning the features of audio signals and then classifying the signals into different categories. Transformers are also well-suited for audio classification. In particular, temporal and frequency features can be extracted to identify the characteristics of the audio signals. Finally, hybrid models for audio classification either combine various deep learning architectures (i.e. CNN-RNN) or combine deep learning models with traditional machine learning techniques (i.e. CNN-Support Vector Machine). These hybrid models take advantage of the strengths of different architectures while avoiding their weaknesses. Existing literature under different categories of deep learning are summarized and compared in detail.