Sivagami Manickam
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Sivagami Manickam has a total of 32 co-authors. Here are the top co-authors by number of shared publications:
- P. Pankaja Lakshmi
Vellore Institute of Technology
3 shared publications - S. Angelin Beulah
School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, India
3 shared publications Angelin Beulah S
School of Computer Science and Engineering, Vellore Institute of Technology
2 shared publications- Jeganathan Lakshmanan
Vellore Institute of Technology University
2 shared publications - Nachimuthu Maheswari
School of Computer Science and Engineering, Vellore Institute of Technology
2 shared publications - Aditi Jain
School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, India
1 shared publication
Latest Publications
- Open Access
- 26 September 2026
- Published by Springer Nature in Scientific Reports
An imbalance-aware class weighted TernausResnet architecture for high-fidelity medical image classification
CWTRNet: A Class weighted TernausResnet framework with adaptive optimization is proposed in this work for high reliability medical image classification. Deep learning(DL) models have made significant advancements in the analysis of medical images, especially in diagnosis and classification. TernausNet, on the other hand, is mostly intended for semantic segmentation rather than classification, yet it successfully maintains multi-scale spatial information using encoder–decoder skip connections. This research gap motivates the creation of a TernausNet adaptation focused on classification that takes the benefit of multi-scale feature refinement without requiring dense segmentation outputs. In order to address class imbalance in image classification, the customized TernausResnet and class-weighting is used in this experimental study. Additionally, training incorporates class-weighted learning to increase minority-class recognition and lessen the impact of class imbalance. The pretrained Resnet-50 backbone offers strong feature representations, which are fine-tuned on medical datasets to enhance diagnostic accuracy. The proposed model was assessed using the F1-score, Geometric Mean, Matthews Correlation Coefficient (MCC), Cohen’s Kappa (CK), and Area Under the Curve (AUC) on chest X-ray dataset under identical learning configurations. The model was evaluated with various medical images such as PneumoniaMNIST and OrganAMNIST and the experimental findings indicate that the CWTRNet achieves high classification accuracy and robust generalization. The proposed model not only enhances classification performance but also improves interpretability by highlighting the key areas affecting the model predictions through Grad-CAM visualization. The model achieves 98.8% training accuracy with a test accuracy of 98.74% and a test loss of 0.0418. The proposed model demonstrated good generalization and low prediction error. Strong discriminative capacity is shown in the classification report for all the classes. The proposed model has been evaluated with MCC and CK metrics for demonstrating per class accuracy metrics. The suggested model performs well per class, as seen by its strong agreement with Cohen’s Kappa = 0.98 and Matthew’s Correlation Coefficient = 0.98. These findings demonstrate that a lightweight classification module for the segmentation output head, combining a residual encoder with decoder-based multi-scale feature refinement class-weighted optimization offers an efficient framework for managing imbalanced medical image classification tasks.
- 26 September 2026
- Published by Springer Nature in Scientific Reports
Open AccessAn imbalance-aware class weighted TernausResnet architecture for high-fidelity medical image classification
CWTRNet: A Class weighted TernausResnet framework with adaptive optimization is proposed in this work for high reliability medical image classification. Deep learning(DL) models have made significant advancements in the analysis of medical images, especially in diagnosis and classification. TernausNet, on the other hand, is mostly intended for semantic segmentation rather than classification, yet it successfully maintains multi-scale spatial information using encoder–decoder skip connections. This research gap motivates the creation of a TernausNet adaptation focused on classification that takes the benefit of multi-scale feature refinement without requiring dense segmentation outputs. In order to address class imbalance in image classification, the customized TernausResnet and class-weighting is used in this experimental study. Additionally, training incorporates class-weighted learning to increase minority-class recognition and lessen the impact of class imbalance. The pretrained Resnet-50 backbone offers strong feature representations, which are fine-tuned on medical datasets to enhance diagnostic accuracy. The proposed model was assessed using the F1-score, Geometric Mean, Matthews Correlation Coefficient (MCC), Cohen’s Kappa (CK), and Area Under the Curve (AUC) on chest X-ray dataset under identical learning configurations. The model was evaluated with various medical images such as PneumoniaMNIST and OrganAMNIST and the experimental findings indicate that the CWTRNet achieves high classification accuracy and robust generalization. The proposed model not only enhances classification performance but also improves interpretability by highlighting the key areas affecting the model predictions through Grad-CAM visualization. The model achieves 98.8% training accuracy with a test accuracy of 98.74% and a test loss of 0.0418. The proposed model demonstrated good generalization and low prediction error. Strong discriminative capacity is shown in the classification report for all the classes. The proposed model has been evaluated with MCC and CK metrics for demonstrating per class accuracy metrics. The suggested model performs well per class, as seen by its strong agreement with Cohen’s Kappa = 0.98 and Matthew’s Correlation Coefficient = 0.98. These findings demonstrate that a lightweight classification module for the segmentation output head, combining a residual encoder with decoder-based multi-scale feature refinement class-weighted optimization offers an efficient framework for managing imbalanced medical image classification tasks.
- Open Access
- 7 September 2026
- Published by Frontiers Media SA in Frontiers in Marine Science
- Vol. 13, 1929977
RSADAF: a Rayleigh scattering-adaptive anisotropic diffusion filter for underwater image restoration in marine ecosystems
Underwater ecosystems play a crucial role in maintaining marine biodiversity and global environmental sustainability. However, underwater monitoring and imaging is a highly challenging area of image processing, where images are severely degraded by haze, wavelength-dependent light absorption, color distortion, scattering, and noise, affecting marine monitoring, object detection, underwater navigation, and limited assessment of coral reefs, marine organisms, benthic habitats, and water quality. The conventional anisotropic diffusion filters do not effectively suppress noise and often fail to adequately preserve image edges and structural details. The manual selection of conductance parameters done in existing methods does not adequately account for the physical degradation mechanisms present in underwater environments. To address these limitations, a new framework, known as Rayleigh scattering-adapted diffusion anisotropic filter (RSADAF), has been proposed, which is a wavelength-aware physics-based diffusion framework that integrates a second-order partial differential equation (PDE) diffusion with the atmospheric scattering model (ASM) and Rayleigh scattering law. The adaptive conductance gradient parameter has been derived from ambient light estimation, and the dominant wavelength of the images has been derived from the CIE 1931 chromaticity model, giving spatial variation of diffusion and reducing local degradations. The second-order PDE formulation (two-stencil approach) preserves the structural information and mitigates staircase artifacts commonly observed in conventional diffusion schemes. Experimental evaluation on the EUVP and UIEB benchmark datasets demonstrates that the proposed approach achieves superior enhancement performance in terms of both full-reference and no-reference image quality metrics. The CIE 1931 wavelength-aware framework achieved a PSNR of 42.85 dB, SSIM of 0.967, UCIQE of 0.341, and BRISQUE of 21.99, outperforming conventional anisotropic diffusion and related enhancement methods. Visual results further confirm improved color restoration, contrast enhancement, edge preservation, and texture retention, offering an interpretable, computationally efficient, and physically meaningful alternative solution for underwater imaging applications.
- 7 September 2026
- Published by Frontiers Media SA in Frontiers in Marine Science
- Vol. 13, 1929977
Open AccessRSADAF: a Rayleigh scattering-adaptive anisotropic diffusion filter for underwater image restoration in marine ecosystems
Underwater ecosystems play a crucial role in maintaining marine biodiversity and global environmental sustainability. However, underwater monitoring and imaging is a highly challenging area of image processing, where images are severely degraded by haze, wavelength-dependent light absorption, color distortion, scattering, and noise, affecting marine monitoring, object detection, underwater navigation, and limited assessment of coral reefs, marine organisms, benthic habitats, and water quality. The conventional anisotropic diffusion filters do not effectively suppress noise and often fail to adequately preserve image edges and structural details. The manual selection of conductance parameters done in existing methods does not adequately account for the physical degradation mechanisms present in underwater environments. To address these limitations, a new framework, known as Rayleigh scattering-adapted diffusion anisotropic filter (RSADAF), has been proposed, which is a wavelength-aware physics-based diffusion framework that integrates a second-order partial differential equation (PDE) diffusion with the atmospheric scattering model (ASM) and Rayleigh scattering law. The adaptive conductance gradient parameter has been derived from ambient light estimation, and the dominant wavelength of the images has been derived from the CIE 1931 chromaticity model, giving spatial variation of diffusion and reducing local degradations. The second-order PDE formulation (two-stencil approach) preserves the structural information and mitigates staircase artifacts commonly observed in conventional diffusion schemes. Experimental evaluation on the EUVP and UIEB benchmark datasets demonstrates that the proposed approach achieves superior enhancement performance in terms of both full-reference and no-reference image quality metrics. The CIE 1931 wavelength-aware framework achieved a PSNR of 42.85 dB, SSIM of 0.967, UCIQE of 0.341, and BRISQUE of 21.99, outperforming conventional anisotropic diffusion and related enhancement methods. Visual results further confirm improved color restoration, contrast enhancement, edge preservation, and texture retention, offering an interpretable, computationally efficient, and physically meaningful alternative solution for underwater imaging applications.
- Open Access
- 18 May 2026
- Published by Frontiers Media SA in Frontiers in Artificial Intelligence
- Vol. 9, 1748985
Artificial intelligence techniques for classification of Alzheimer's disease using neuroimaging data: a review
Alzheimer's disease (AD) is a gradually advancing brain disorder marked by memory impairment. The incurable, progressive nature of the disease leads to the dementia stage. Treatment is effective in the early stage, and it can be controlled but not cured. Artificial Intelligence ( AI) learning models are used in medical science to detect and classify diseases into specific categories. Features are extracted from medical images and trained using AI learning models to perform an accurate diagnosis of AD. Recent advancements in machine learning (ML) and deep learning (DL) models have demonstrated significant potential in identifying AD across various data modalities, including neuroimaging, genetic information, and clinical assessments. This study focuses on the application of advanced ML and DL techniques in the identification and classification of AD, including regression models, decision trees, random forests, support vector machines (SVMs), k-nearest neighbors (KNNs), ensemble models, convolutional neural networks (CNNs), recurrent neural networks (RNNs), and generative adversarial networks (GANs). Each model is analyzed for its strengths, limitations, and performance metrics, with particular emphasis on the importance of data preprocessing and augmentation techniques to improve model accuracy and robustness. The review highlights that multimodal approaches, particularly the fusion of MRI and PET data, enhance classification accuracy compared to single-modality models. Additionally, transfer learning techniques have shown promise in overcoming data limitations by leveraging pretrained models. The review also highlights the critical role of evaluation metrics in assessing model performance, emphasizing the need for a diverse set that includes accuracy, precision, recall, F1-score, and Cohen's Kappa. The study identifies gaps in the current literature, including underreporting of certain metrics and the need for more comprehensive evaluations, and provides recommendations for future research. Finally, this study discusses the challenges and opportunities in the field, including improving model generalizability, enhancing interpretability, advanced data preprocessing and augmentation, integration with clinical workflows, and multimodal data fusion. This review provides consolidated information that may be useful for researchers, clinicians, and data scientists, offering insights into current trends, challenges, and future research directions in AI-driven AD detection.
- 18 May 2026
- Published by Frontiers Media SA in Frontiers in Artificial Intelligence
- Vol. 9, 1748985
Open AccessArtificial intelligence techniques for classification of Alzheimer's disease using neuroimaging data: a review
Alzheimer's disease (AD) is a gradually advancing brain disorder marked by memory impairment. The incurable, progressive nature of the disease leads to the dementia stage. Treatment is effective in the early stage, and it can be controlled but not cured. Artificial Intelligence ( AI) learning models are used in medical science to detect and classify diseases into specific categories. Features are extracted from medical images and trained using AI learning models to perform an accurate diagnosis of AD. Recent advancements in machine learning (ML) and deep learning (DL) models have demonstrated significant potential in identifying AD across various data modalities, including neuroimaging, genetic information, and clinical assessments. This study focuses on the application of advanced ML and DL techniques in the identification and classification of AD, including regression models, decision trees, random forests, support vector machines (SVMs), k-nearest neighbors (KNNs), ensemble models, convolutional neural networks (CNNs), recurrent neural networks (RNNs), and generative adversarial networks (GANs). Each model is analyzed for its strengths, limitations, and performance metrics, with particular emphasis on the importance of data preprocessing and augmentation techniques to improve model accuracy and robustness. The review highlights that multimodal approaches, particularly the fusion of MRI and PET data, enhance classification accuracy compared to single-modality models. Additionally, transfer learning techniques have shown promise in overcoming data limitations by leveraging pretrained models. The review also highlights the critical role of evaluation metrics in assessing model performance, emphasizing the need for a diverse set that includes accuracy, precision, recall, F1-score, and Cohen's Kappa. The study identifies gaps in the current literature, including underreporting of certain metrics and the need for more comprehensive evaluations, and provides recommendations for future research. Finally, this study discusses the challenges and opportunities in the field, including improving model generalizability, enhancing interpretability, advanced data preprocessing and augmentation, integration with clinical workflows, and multimodal data fusion. This review provides consolidated information that may be useful for researchers, clinicians, and data scientists, offering insights into current trends, challenges, and future research directions in AI-driven AD detection.
- Open Access
- 4 May 2026
- Published by Elsevier in SSRN Electronic Journal
A Novel Dynamic Quantum-Inspired Attention Mechanism with Residual Feature Fusion for Medical Image Classification
Medical image classification plays a vital role in computer-aided diagnosis. However, its performance is often limited by high-dimensional feature representations, data scarcity, and the inability of classical models to effectively capture complex feature interactions. In this paper, we propose an enhanced framework integrating Adaptive Entangled Quantum Fourier Feature Learning (AEQFFL) with a novel Dynamic Quantum-Attention Module (DQAM) to improve classification performance. The proposed approach transforms high-dimensional representations into a compact quantum-compatible space using a learnable feature compression mechanism after robust feature extraction through a customized deep convolutional neural network based on ResNetwithGoldenRatio-basedactivation (GldReLU). Global dependencies and phase correla tions between features are captured using an encoding based on the Quantum Fourier Transform (QFT) with parameterized rotations and entanglement operations. The proposed DQAMNet combines Dynamic Feature Selection (DFS) with quantum-inspired attention to further improve feature representation. This enables adaptive weighting of important features while suppressing irrelevant information. Stable learning and efficient retention of original features are guaranteed by the integration of residual fusion. To register for the benefit of contrasting information, a hybrid fusion method integrates classical and quantum representations. The advantage of merging quantum-inspired learning with advanced attention mechanisms for medical image analysis is shown by experimental results, which demonstrate that the suggested framework significantly improves classification performance and feature discriminability, especially in instances with limited labeled data and complex feature distributions
- 4 May 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA Novel Dynamic Quantum-Inspired Attention Mechanism with Residual Feature Fusion for Medical Image Classification
Medical image classification plays a vital role in computer-aided diagnosis. However, its performance is often limited by high-dimensional feature representations, data scarcity, and the inability of classical models to effectively capture complex feature interactions. In this paper, we propose an enhanced framework integrating Adaptive Entangled Quantum Fourier Feature Learning (AEQFFL) with a novel Dynamic Quantum-Attention Module (DQAM) to improve classification performance. The proposed approach transforms high-dimensional representations into a compact quantum-compatible space using a learnable feature compression mechanism after robust feature extraction through a customized deep convolutional neural network based on ResNetwithGoldenRatio-basedactivation (GldReLU). Global dependencies and phase correla tions between features are captured using an encoding based on the Quantum Fourier Transform (QFT) with parameterized rotations and entanglement operations. The proposed DQAMNet combines Dynamic Feature Selection (DFS) with quantum-inspired attention to further improve feature representation. This enables adaptive weighting of important features while suppressing irrelevant information. Stable learning and efficient retention of original features are guaranteed by the integration of residual fusion. To register for the benefit of contrasting information, a hybrid fusion method integrates classical and quantum representations. The advantage of merging quantum-inspired learning with advanced attention mechanisms for medical image analysis is shown by experimental results, which demonstrate that the suggested framework significantly improves classification performance and feature discriminability, especially in instances with limited labeled data and complex feature distributions
- Open Access
- 16 March 2026
- Published by Elsevier in SSRN Electronic Journal
Adaptive Edge-Aware Denoising Framework with Quantum Analysis for Ultrasound ImageEnhancement
Ultrasound imaging is widely used for medical diagnosis; however, speckle noise and low contrast degrade image quality and diagnostic reliability. Conventional denoising approaches are based on filtering and transform-domain techniques, suffering from information loss, leading to blurred edges and degraded fine structures. To address these limitations, a hybrid ultrasound image enhancement framework that integrates an edge-constrained denoising autoencoder (E-DAE) with a cascaded adaptive convolutional neural network (CACNN) is developed. Ultrasound images were preprocessed via gray-world normalization and contrast-limited adaptive histogram equalization (CLAHE) to improve illumination consistency and local contrast. Speckle noise is then suppressed with edge-constrained denoising autoencoder (E-DAE), which employs dual-domain loss function that combines pixel-level reconstruction loss and gradient-domain edge preservation to effectively reduce noise while maintaining critical structural boundaries. The denoised images are subsequently refined through a CACNN, where progressive feature learning with ReLU activations enables high-resolution enhancement and fine-detail preservation. Each cascaded stage consists of 20 convolutional layers, and experimental analysis reveals that network performance saturates beyond 40 layers, indicating that no further benefit from increased depth. Compared with recent state-of-the-art methods, comprehensive quantitative and qualitative evaluations demonstrated significant improvements in image quality metrics, including PSNR, SSIM, normalized absolute error (NAE), the speckle suppression index (SSI), the contrast-to-noise ratio (CNR), and the equivalent number of look (ENL). Furthermore, quantum convolutional neural networks (QCNNs) were explored for comparative analysis. Although the QCNNs exhibited a reduced execution time, they failed to achieve comparable improvements in either quantitative or qualitative image quality metrics relative to the proposed CACNN-based framework.
- 16 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessAdaptive Edge-Aware Denoising Framework with Quantum Analysis for Ultrasound ImageEnhancement
Ultrasound imaging is widely used for medical diagnosis; however, speckle noise and low contrast degrade image quality and diagnostic reliability. Conventional denoising approaches are based on filtering and transform-domain techniques, suffering from information loss, leading to blurred edges and degraded fine structures. To address these limitations, a hybrid ultrasound image enhancement framework that integrates an edge-constrained denoising autoencoder (E-DAE) with a cascaded adaptive convolutional neural network (CACNN) is developed. Ultrasound images were preprocessed via gray-world normalization and contrast-limited adaptive histogram equalization (CLAHE) to improve illumination consistency and local contrast. Speckle noise is then suppressed with edge-constrained denoising autoencoder (E-DAE), which employs dual-domain loss function that combines pixel-level reconstruction loss and gradient-domain edge preservation to effectively reduce noise while maintaining critical structural boundaries. The denoised images are subsequently refined through a CACNN, where progressive feature learning with ReLU activations enables high-resolution enhancement and fine-detail preservation. Each cascaded stage consists of 20 convolutional layers, and experimental analysis reveals that network performance saturates beyond 40 layers, indicating that no further benefit from increased depth. Compared with recent state-of-the-art methods, comprehensive quantitative and qualitative evaluations demonstrated significant improvements in image quality metrics, including PSNR, SSIM, normalized absolute error (NAE), the speckle suppression index (SSI), the contrast-to-noise ratio (CNR), and the equivalent number of look (ENL). Furthermore, quantum convolutional neural networks (QCNNs) were explored for comparative analysis. Although the QCNNs exhibited a reduced execution time, they failed to achieve comparable improvements in either quantitative or qualitative image quality metrics relative to the proposed CACNN-based framework.
- Open Access
- 26 February 2026
- Published by Springer Nature in Discover Computing
- Vol. 29 (1)
Color correction model for underwater images using a lightweight truncated U-Net
- 26 February 2026
- Published by Springer Nature in Discover Computing
- Vol. 29 (1)
Open AccessColor correction model for underwater images using a lightweight truncated U-Net
- 4 August 2025
- Published by Springer Nature in Neural Computing and Applications
- Vol. 37 (26) , 21997-22028
A novel GldReLU activation function with enhanced RESNET50 for classification of X-ray images
- 4 August 2025
- Published by Springer Nature in Neural Computing and Applications
- Vol. 37 (26) , 21997-22028
A novel GldReLU activation function with enhanced RESNET50 for classification of X-ray images
- Open Access
- 22 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 149637-149656
UResNet-Based Enhancement of Underwater Images Through Variational Contrast and Saturation
Underwater imaging is an ever trending and evolving field in which the images obtained are subjected to enhancement to identify the critical features of the underwater world. Various enhancement approaches have evolved and certain issues have been identified in these techniques leading to a new technique. Generally, underwater images suffer from blurriness, color distortion, light absorption and scattering in water, low contrast issues which affect the overall clarity and visual appeal of the image. Our research demonstrates that hybrid preprocessing when combined with advanced deep learning models such as Underwater ResNet (UResNet) give high quality and visually appealing images. Three model variants have been explored in this work, the standard UResNet, UResNet augmented with a Sobel filter, and UResNet enhanced with a squeeze-and-excitation (SE) block. The models have employed on the EUVP (Enhancing Underwater Visual Perception) dataset. The core challenges such as color distortion, low contrast, and poor saturation have been overcome by the hybrid preprocessing pipeline, which combines Underwater White Balance (UWB) and Variational Contrast and Saturation Enhancement (VCSE) techniques that corrects the color imbalances, enhances the contrast, and amplifies the saturation before deep learning interference. The pre-processed images were given to all modified UResNet models and it was found that all models outperform the baseline model in both qualitative and quantitative evaluations. Among the three models that were modified, the SE-integrated model consistently delivered superior performance with all the metrics, highlighting significant improvements in color fidelity, contrast, edge definition, and overall image clarity. These results show the effectiveness of combining hybrid preprocessing with structure-aware deep learning architectures to improve underwater image quality and adaptability across diverse marine environments.
- 22 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 149637-149656
Open AccessUResNet-Based Enhancement of Underwater Images Through Variational Contrast and Saturation
Underwater imaging is an ever trending and evolving field in which the images obtained are subjected to enhancement to identify the critical features of the underwater world. Various enhancement approaches have evolved and certain issues have been identified in these techniques leading to a new technique. Generally, underwater images suffer from blurriness, color distortion, light absorption and scattering in water, low contrast issues which affect the overall clarity and visual appeal of the image. Our research demonstrates that hybrid preprocessing when combined with advanced deep learning models such as Underwater ResNet (UResNet) give high quality and visually appealing images. Three model variants have been explored in this work, the standard UResNet, UResNet augmented with a Sobel filter, and UResNet enhanced with a squeeze-and-excitation (SE) block. The models have employed on the EUVP (Enhancing Underwater Visual Perception) dataset. The core challenges such as color distortion, low contrast, and poor saturation have been overcome by the hybrid preprocessing pipeline, which combines Underwater White Balance (UWB) and Variational Contrast and Saturation Enhancement (VCSE) techniques that corrects the color imbalances, enhances the contrast, and amplifies the saturation before deep learning interference. The pre-processed images were given to all modified UResNet models and it was found that all models outperform the baseline model in both qualitative and quantitative evaluations. Among the three models that were modified, the SE-integrated model consistently delivered superior performance with all the metrics, highlighting significant improvements in color fidelity, contrast, edge definition, and overall image clarity. These results show the effectiveness of combining hybrid preprocessing with structure-aware deep learning architectures to improve underwater image quality and adaptability across diverse marine environments.
- Open Access
- 9 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 10582-10594
Enhanced YOLOv8 Object Detection Model for Construction Worker Safety Using Image Transformations
The rapid growth of Deep Learning techniques plays a vital role in automation of manual work in various areas. One such area for application of new technology is that of Construction Worker Safety. It has thus become imperative to improve existing systems with the new capabilities of technology. This paper discusses a methodology of improving the performance of an existing approach of object detection, YOLOv8. The proposed work comprises of improved training of model and detection of helmet in worker images, using Test Time Augmentation (TTA) based approach. Image Transformations such as Histogram Equalization, Gamma Correction, Gaussian Blurring and Contrast Stretching are applied to augment the dataset by creating more versions of the existing data. This has shown to improve the performance of the model and also generalize better by preventing overfitting. A Test Time Augmentation-based Confidence Thresholding formula (TTACT) is also proposed, to improve the performance of helmet detection.
- 9 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 10582-10594
Open AccessEnhanced YOLOv8 Object Detection Model for Construction Worker Safety Using Image Transformations
The rapid growth of Deep Learning techniques plays a vital role in automation of manual work in various areas. One such area for application of new technology is that of Construction Worker Safety. It has thus become imperative to improve existing systems with the new capabilities of technology. This paper discusses a methodology of improving the performance of an existing approach of object detection, YOLOv8. The proposed work comprises of improved training of model and detection of helmet in worker images, using Test Time Augmentation (TTA) based approach. Image Transformations such as Histogram Equalization, Gamma Correction, Gaussian Blurring and Contrast Stretching are applied to augment the dataset by creating more versions of the existing data. This has shown to improve the performance of the model and also generalize better by preventing overfitting. A Test Time Augmentation-based Confidence Thresholding formula (TTACT) is also proposed, to improve the performance of helmet detection.
- 3 November 2024
- Published by Springer Nature
- p. 300-312
Securing Data in MongoDB: A Framework Using Encryption
- 3 November 2024
- Published by Springer Nature
- p. 300-312
Securing Data in MongoDB: A Framework Using Encryption