Kuntao Ye
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Kuntao Ye has a total of 24 co-authors. Here are the top co-authors by number of shared publications:
- Zhiqiang Xiao
School of Science, Jiangxi University of Science and Technology, Ganzhou, China
3 shared publications - Guolong Cui
University of Electronic Science and Technology of China
2 shared publications Shubin Qiu
School of Science, Jiangxi University of Science and Technology, Ganzhou, China
2 shared publications- Yue Zhang
School of Science, Jiangxi University of Science and Technology, Ganzhou, Jiangxi, China
2 shared publications - Chunju Hou
School of Science, Jiangxi University of Science and Technology, Ganzhou, 341000, China
1 shared publication Dong Tai-Yuan
1 shared publication
Latest Publications
- Open Access
- 21 March 2025
- Published by Association for Computing Machinery (ACM)
- p. 282-286
Deep learning optimizer based on metaheuristic algorithms
- 21 March 2025
- Published by Association for Computing Machinery (ACM)
- p. 282-286
Open AccessDeep learning optimizer based on metaheuristic algorithms
- Open Access
- 5 March 2025
- Published by Springer Nature in Scientific Reports
- Vol. 15 (1) , 1-19
Study on lightweight strategies for L-YOLO algorithm in road object detection
With the increasing complexity of urban traffic, object detection has become critical in autonomous driving and intelligent traffic management. The demand for real-time, efficient object detection systems is growing. However, traditional algorithms often suffer from large parameter sizes and high computational costs, limiting their applicability in resource-constrained environments. To address this issue, we propose L-YOLO, an improved lightweight road object detection algorithm based on YOLOv8s. First, L-HGNetV2 replaces the backbone network of YOLOv8s to enhance feature extraction and fusion efficiency. Second, a small object detection layer is introduced into the feature fusion network, replacing the original C2f modules with the new CStar modules. This modification improves the capture of features and contextual information for small vehicle targets without significantly increasing computational demands. Third, the CIoU loss function is replaced by the FPIoU2 loss function, enhancing the model’s robustness. Finally, the layer adaptive magnitude-based model pruning (LAMP) method is applied to prune the convolutional layer channels, significantly reducing the computational burden and parameter count while maintaining accuracy, thus improving operational efficiency. On the KITTI public dataset, L-YOLO achieves a mAP50 of 93.8%, a 2.5% improvement over YOLOv8s. The number of parameters decreases from 11.12 M to 3.58 M, and the computational load is reduced from 28.4 GFLOPs to 14.2 GFLOPs.
- 5 March 2025
- Published by Springer Nature in Scientific Reports
- Vol. 15 (1) , 1-19
Open AccessStudy on lightweight strategies for L-YOLO algorithm in road object detection
With the increasing complexity of urban traffic, object detection has become critical in autonomous driving and intelligent traffic management. The demand for real-time, efficient object detection systems is growing. However, traditional algorithms often suffer from large parameter sizes and high computational costs, limiting their applicability in resource-constrained environments. To address this issue, we propose L-YOLO, an improved lightweight road object detection algorithm based on YOLOv8s. First, L-HGNetV2 replaces the backbone network of YOLOv8s to enhance feature extraction and fusion efficiency. Second, a small object detection layer is introduced into the feature fusion network, replacing the original C2f modules with the new CStar modules. This modification improves the capture of features and contextual information for small vehicle targets without significantly increasing computational demands. Third, the CIoU loss function is replaced by the FPIoU2 loss function, enhancing the model’s robustness. Finally, the layer adaptive magnitude-based model pruning (LAMP) method is applied to prune the convolutional layer channels, significantly reducing the computational burden and parameter count while maintaining accuracy, thus improving operational efficiency. On the KITTI public dataset, L-YOLO achieves a mAP50 of 93.8%, a 2.5% improvement over YOLOv8s. The number of parameters decreases from 11.12 M to 3.58 M, and the computational load is reduced from 28.4 GFLOPs to 14.2 GFLOPs.
- Open Access
- 9 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 9524-9537
Multi-Scale Channel Distillation Network for Image Compressive Sensing
Recently, convolutional neural networks (CNNs) have demonstrated striking success in computer vision tasks. Methods based on CNNs for image compressive sensing (CS) have also gained prominence. However, existing methods tend to increase the depth of the network in feature space for better reconstruction quality, neglecting the hierarchical representation of intermediate features in pixel space. In order to coordinate the feature space and pixel space to complete the deep reconstruction of images, and further improve the reconstruction performance of current CS methods, we propose a Multi-Scale Channel Distillation Network (MSCDN). This network first decomposes an input image in scale space at the sampling stage, followed by sampling these decomposed images through a convolutional operation. In this way, multiscale information of the input in the compressed domain is aggregated. During the reconstruction phase, a low-frequency information recovery network generates a preliminary image, whereas a high-frequency feature aggregation network refines the image further. Specifically, we design a dual-branch deep reconstruction architecture with Channel Distillation Residual Block (CDRB) as the core component. One branch extracts features gradually by cascading multiple CDRB modules, thereby supplementing the initial reconstructed image with a large amount of high-frequency content in feature space. The other branch takes the initial reconstructed image as input and sequentially fuses the intermediate feature outputs by CDRBs to enhance the local details of the image in pixel space. Combining outputs from both branches, we achieve an optimal reconstructed image. Experimental results on four test datasets demonstrate that MSCDN surpasses state-of-the-art CS methods not only in reconstruction accuracy but also in perceptual visual quality.
- 9 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 9524-9537
Open AccessMulti-Scale Channel Distillation Network for Image Compressive Sensing
Recently, convolutional neural networks (CNNs) have demonstrated striking success in computer vision tasks. Methods based on CNNs for image compressive sensing (CS) have also gained prominence. However, existing methods tend to increase the depth of the network in feature space for better reconstruction quality, neglecting the hierarchical representation of intermediate features in pixel space. In order to coordinate the feature space and pixel space to complete the deep reconstruction of images, and further improve the reconstruction performance of current CS methods, we propose a Multi-Scale Channel Distillation Network (MSCDN). This network first decomposes an input image in scale space at the sampling stage, followed by sampling these decomposed images through a convolutional operation. In this way, multiscale information of the input in the compressed domain is aggregated. During the reconstruction phase, a low-frequency information recovery network generates a preliminary image, whereas a high-frequency feature aggregation network refines the image further. Specifically, we design a dual-branch deep reconstruction architecture with Channel Distillation Residual Block (CDRB) as the core component. One branch extracts features gradually by cascading multiple CDRB modules, thereby supplementing the initial reconstructed image with a large amount of high-frequency content in feature space. The other branch takes the initial reconstructed image as input and sequentially fuses the intermediate feature outputs by CDRBs to enhance the local details of the image in pixel space. Combining outputs from both branches, we achieve an optimal reconstructed image. Experimental results on four test datasets demonstrate that MSCDN surpasses state-of-the-art CS methods not only in reconstruction accuracy but also in perceptual visual quality.
- 1 October 2024
- Published by SPIE-Intl Soc Optical Eng in Journal of Applied Remote Sensing
- Vol. 18 (04) , 046506-15
Lightweight multi-scale bidirectional network for remote sensing image dehazing
Many existing multi-scale approaches for remote sensing image dehazing typically employ a unidirectional strategy, progressing from coarse to fine scales. This approach often leads to poor dehazing performance because information from subsequent scales is not fully leveraged and estimation errors at coarser scales will propagate to finer scales. To overcome the shortcomings and enhance the performance of image dehazing, bidirectional mechanisms are introduced to our proposed lightweight multi-scale bidirectional network (MSB-Net). The submitted MSB-Net comprises three intra-scale branches and two inter-scale branches. In the intra-scale branches, Transformer-based UNet integrates non-local sparse aggregation blocks to enhance global information recovery and improve modeling efficiency by learning global features at different scales. In the inter-scale branches, convolutional neural network (CNN)-based UNet augmented with bidirectional feature propagation blocks (BFPB) is employed to adaptively aggregate the multi-scale features at different scales. Although the local feature enhancement block in the CNN-based UNet is central to emphasizing local features, BFPB effectively exploits information spanning various scales for reconstruction by allowing bidirectional information propagation in both coarse-to-fine and fine-to-coarse flows. Experimental results demonstrate that MSB-Net outperforms other popular Transformer-based approaches, particularly achieving a 16% reduction in model parameters.
- 1 October 2024
- Published by SPIE-Intl Soc Optical Eng in Journal of Applied Remote Sensing
- Vol. 18 (04) , 046506-15
Lightweight multi-scale bidirectional network for remote sensing image dehazing
Many existing multi-scale approaches for remote sensing image dehazing typically employ a unidirectional strategy, progressing from coarse to fine scales. This approach often leads to poor dehazing performance because information from subsequent scales is not fully leveraged and estimation errors at coarser scales will propagate to finer scales. To overcome the shortcomings and enhance the performance of image dehazing, bidirectional mechanisms are introduced to our proposed lightweight multi-scale bidirectional network (MSB-Net). The submitted MSB-Net comprises three intra-scale branches and two inter-scale branches. In the intra-scale branches, Transformer-based UNet integrates non-local sparse aggregation blocks to enhance global information recovery and improve modeling efficiency by learning global features at different scales. In the inter-scale branches, convolutional neural network (CNN)-based UNet augmented with bidirectional feature propagation blocks (BFPB) is employed to adaptively aggregate the multi-scale features at different scales. Although the local feature enhancement block in the CNN-based UNet is central to emphasizing local features, BFPB effectively exploits information spanning various scales for reconstruction by allowing bidirectional information propagation in both coarse-to-fine and fine-to-coarse flows. Experimental results demonstrate that MSB-Net outperforms other popular Transformer-based approaches, particularly achieving a 16% reduction in model parameters.
- Open Access
- 26 April 2024
- Published by Association for Computing Machinery (ACM)
- p. 135-139
An improved YOLOV5 based water surface garbage detection algorithm
A single-stage surface garbage detection model with higher accuracy is proposed to address the problem of difficulty detecting water surface garbage in complex environments. Firstly, the C3RFE module is designed by combining TridentNet with dilated convolution, which can increase the receptive field of the model and enable the model to obtain more global information. Then, the CBAM attention mechanism is integrated into the decoupling detection head, so that the model classification and regression tasks are carried out separately, reducing the information redundancy and improving the model's adaptive feature refinement ability. Finally, the sigmoid function is used to optimize the penalty factor of the CIOU loss function, which makes the loss function smoother and more robust. Through experiments on the public FIoW-img dataset, the results show that the algorithm is superior to other mainstream detection models, and is 2.5 percent higher than the baseline YOLOV5 s model mAP, 0.5 percent higher than Recall, and 5 percent higher than Precision.
- 26 April 2024
- Published by Association for Computing Machinery (ACM)
- p. 135-139
Open AccessAn improved YOLOV5 based water surface garbage detection algorithm
A single-stage surface garbage detection model with higher accuracy is proposed to address the problem of difficulty detecting water surface garbage in complex environments. Firstly, the C3RFE module is designed by combining TridentNet with dilated convolution, which can increase the receptive field of the model and enable the model to obtain more global information. Then, the CBAM attention mechanism is integrated into the decoupling detection head, so that the model classification and regression tasks are carried out separately, reducing the information redundancy and improving the model's adaptive feature refinement ability. Finally, the sigmoid function is used to optimize the penalty factor of the CIOU loss function, which makes the loss function smoother and more robust. Through experiments on the public FIoW-img dataset, the results show that the algorithm is superior to other mainstream detection models, and is 2.5 percent higher than the baseline YOLOV5 s model mAP, 0.5 percent higher than Recall, and 5 percent higher than Precision.
- 22 February 2024
- Published by World Scientific Pub Co Pte Ltd in Journal of Circuits, Systems and Computers
- Vol. 33 (12)
A Statistical Model-Based Approach for Reproducing Intermittent Faults in Electrical Connectors under Varying Vibration Loading Conditions
The performance of electrical connectors can be significantly impacted by periodic variations in contact resistance caused by vibrational stress. Intermittent faults resulting from such stress are characterized by their random and fleeting nature, making it difficult to study and replicate them. This paper proposes a novel method for reproducing intermittent faults in electrical connectors. To implement this method, intermittent fault data are first collected from electrical connectors subjected to different vibration loads. Next, a statistical distribution model is constructed using kernel density estimation (KDE). Based on this model, a fault injector is designed to simulate intermittent faults under varying vibration loads. The simulated faults are then compared to real-world intermittent fault signals in a controlled environment to validate the accuracy of the method. The results demonstrate that the proposed method effectively reproduces intermittent faults in electrical connectors under varying vibration conditions. This approach can be used to better understand the behavior of connectors under vibrational stress and to develop more effective testing and fault diagnosis methods.
- 22 February 2024
- Published by World Scientific Pub Co Pte Ltd in Journal of Circuits, Systems and Computers
- Vol. 33 (12)
A Statistical Model-Based Approach for Reproducing Intermittent Faults in Electrical Connectors under Varying Vibration Loading Conditions
The performance of electrical connectors can be significantly impacted by periodic variations in contact resistance caused by vibrational stress. Intermittent faults resulting from such stress are characterized by their random and fleeting nature, making it difficult to study and replicate them. This paper proposes a novel method for reproducing intermittent faults in electrical connectors. To implement this method, intermittent fault data are first collected from electrical connectors subjected to different vibration loads. Next, a statistical distribution model is constructed using kernel density estimation (KDE). Based on this model, a fault injector is designed to simulate intermittent faults under varying vibration loads. The simulated faults are then compared to real-world intermittent fault signals in a controlled environment to validate the accuracy of the method. The results demonstrate that the proposed method effectively reproduces intermittent faults in electrical connectors under varying vibration conditions. This approach can be used to better understand the behavior of connectors under vibrational stress and to develop more effective testing and fault diagnosis methods.
- Open Access
- 19 January 2024
- Published by Association for Computing Machinery (ACM)
- p. 624-628
A lightweight transformer based chest X-ray image semantic segmentation model for COVID-19 fast detection
Although convolutional neural network (CNN)-based models can provide high COVID-19 detection accuracy from Chest X-ray (CXR) images, it is still desirable to obtain a fast, lightweight and effective diagnostic deep learning (DL) model. To solve this problem, a Transformer-based semantic segmentation model with an encoding-decoding structure is proposed in this study. In the proposed model, a semantic segmentation structure dubbed as TD-SW-SETR is constructed by introducing tree-structured residual block (TRSB) and shifted-window multi-head self-attention (SW-MSA) mechanism to minimize the number of parameters. The experimental results show that the proposed model requires less trainable parameters to achieve the highest evaluation scores in terms of F1 score (94.2%), Miou (89.6%), and specificity (99.1%) compared to other popular network models.
- 19 January 2024
- Published by Association for Computing Machinery (ACM)
- p. 624-628
Open AccessA lightweight transformer based chest X-ray image semantic segmentation model for COVID-19 fast detection
Although convolutional neural network (CNN)-based models can provide high COVID-19 detection accuracy from Chest X-ray (CXR) images, it is still desirable to obtain a fast, lightweight and effective diagnostic deep learning (DL) model. To solve this problem, a Transformer-based semantic segmentation model with an encoding-decoding structure is proposed in this study. In the proposed model, a semantic segmentation structure dubbed as TD-SW-SETR is constructed by introducing tree-structured residual block (TRSB) and shifted-window multi-head self-attention (SW-MSA) mechanism to minimize the number of parameters. The experimental results show that the proposed model requires less trainable parameters to achieve the highest evaluation scores in terms of F1 score (94.2%), Miou (89.6%), and specificity (99.1%) compared to other popular network models.
- 4 January 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 24 (4) , 5370-5382
PointNet-Transformer Fusion Network for In-Cabin Occupancy Monitoring With mm-Wave Radar
Due to the irregular distribution of 3D point clouds and the random movement of the passengers, it is difficult to accurately and rapidly determine the passenger occupancy based on the multiple-input and multiple-output (MIMO) radar. In this study, we present a lightweight neural network, dubbed PTFNet, that fuses the PointNet and Transformer to quickly and accurately detect the occupancy of three rear seats and two footwells. PTFNet adapts an input transform block to encode the 3D point cloud directly, which makes it more lightweight than conventional voxelization, heatmap, and PointNet based algorithms. Meanwhile, a cross-attention mechanism block inspired by Transformer is employed to efficiently extract features from 3D point clouds, resulting in an improvement in detection accuracy. An in-cabin occupancy monitoring system is implemented in a real vehicle to obtain 3D point cloud datasets, which are then used to evaluate the detection performance of PTFNet. Our datasets cover both the rear seat area and the footwell area. The footwell space is considered for the first time in the point cloud datasets. The experimental results demonstrate that PTFNet outperforms other popular approaches in terms of detection accuracy by at least 1.6%, while the consumption of memory and running time are reduced by at least 44% and 34%, respectively.
- 4 January 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 24 (4) , 5370-5382
PointNet-Transformer Fusion Network for In-Cabin Occupancy Monitoring With mm-Wave Radar
Due to the irregular distribution of 3D point clouds and the random movement of the passengers, it is difficult to accurately and rapidly determine the passenger occupancy based on the multiple-input and multiple-output (MIMO) radar. In this study, we present a lightweight neural network, dubbed PTFNet, that fuses the PointNet and Transformer to quickly and accurately detect the occupancy of three rear seats and two footwells. PTFNet adapts an input transform block to encode the 3D point cloud directly, which makes it more lightweight than conventional voxelization, heatmap, and PointNet based algorithms. Meanwhile, a cross-attention mechanism block inspired by Transformer is employed to efficiently extract features from 3D point clouds, resulting in an improvement in detection accuracy. An in-cabin occupancy monitoring system is implemented in a real vehicle to obtain 3D point cloud datasets, which are then used to evaluate the detection performance of PTFNet. Our datasets cover both the rear seat area and the footwell area. The footwell space is considered for the first time in the point cloud datasets. The experimental results demonstrate that PTFNet outperforms other popular approaches in terms of detection accuracy by at least 1.6%, while the consumption of memory and running time are reduced by at least 44% and 34%, respectively.
- 2 January 2024
- Published by Springer Nature in Soft Computing
- Vol. 28 (9-10) , 6555-6570
An improved beetle swarm antennae search algorithm based on multiple operators
- 2 January 2024
- Published by Springer Nature in Soft Computing
- Vol. 28 (9-10) , 6555-6570
An improved beetle swarm antennae search algorithm based on multiple operators
- 13 December 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 73 (00189456) , 1-17
Differential Self-Feedback Dilated Convolution Network With Dual-Tree Channel Attention Mechanism for Hyperspectral Image Classification
Overfitting and inadequate ability of feature extraction result in low classification accuracy when fulfilling hyperspectral image (HSI) classification tasks by deep learning models (DLMs). Unfortunately, it is very common that insufficient labeled training samples of HSI lead to overfitting and that redundant spectral information of HSI asks for high feature extraction ability in classification tasks. In general, reducing model complexity helps alleviate overfitting, whereas increasing the number of convolutional layers and kernel size helps enhance feature extraction ability. Designing DLMs for better classification accuracy requires balancing model complexity and feature extraction ability. To achieve this balance, we present a network model of differential self-feedback dilated convolution with dual-tree channel attention mechanism (DSDC-DCAM) model. In this model, a differential self-feedback dilated convolution (DSDC) block is proposed with high extraction ability and low model complexity by connecting 3-D dilated convolutions and the differential self-feedback (DSF) module as a tree structure. A dual-tree channel attention mechanism (DCAM) block is also proposed in DSDC-DCAM to form a better attention mechanism in order to exploit useful features and suppress unfavorable ones while reducing model complexity. Experimental results show that DSDC-DCAM has the best performance consistency compared with the other nine state-of-the-art models, indicating that it is least influenced by overfitting. Meanwhile, the results show that DSDC-DCAM achieves the highest overall accuracy (OA) with a smaller number of model parameters than the other seven models.
- 13 December 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 73 (00189456) , 1-17
Differential Self-Feedback Dilated Convolution Network With Dual-Tree Channel Attention Mechanism for Hyperspectral Image Classification
Overfitting and inadequate ability of feature extraction result in low classification accuracy when fulfilling hyperspectral image (HSI) classification tasks by deep learning models (DLMs). Unfortunately, it is very common that insufficient labeled training samples of HSI lead to overfitting and that redundant spectral information of HSI asks for high feature extraction ability in classification tasks. In general, reducing model complexity helps alleviate overfitting, whereas increasing the number of convolutional layers and kernel size helps enhance feature extraction ability. Designing DLMs for better classification accuracy requires balancing model complexity and feature extraction ability. To achieve this balance, we present a network model of differential self-feedback dilated convolution with dual-tree channel attention mechanism (DSDC-DCAM) model. In this model, a differential self-feedback dilated convolution (DSDC) block is proposed with high extraction ability and low model complexity by connecting 3-D dilated convolutions and the differential self-feedback (DSF) module as a tree structure. A dual-tree channel attention mechanism (DCAM) block is also proposed in DSDC-DCAM to form a better attention mechanism in order to exploit useful features and suppress unfavorable ones while reducing model complexity. Experimental results show that DSDC-DCAM has the best performance consistency compared with the other nine state-of-the-art models, indicating that it is least influenced by overfitting. Meanwhile, the results show that DSDC-DCAM achieves the highest overall accuracy (OA) with a smaller number of model parameters than the other seven models.