Dejene Mengistu Sime
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Dejene Mengistu Sime has a total of 20 co-authors. Here are the top co-authors by number of shared publications:
- Kai Sheng
Guangzhou Institute of Technology, Xidian University, Guangzhou, China
4 shared publications - Guotai Wang
University of Electronic Science and Technology of China
3 shared publications - Nan Ouyang
Guangzhou Institute of Technology, Xidian University, Guangzhou, China
3 shared publications - Xiaojiang Ren
Guangzhou Institute of Technology, Xidian University, Guangzhou, China
3 shared publications - Adnan A. Qaseem
School of Electronic Engineering, Xidian University, Xian, China
2 shared publications - Bei Peng
School of Computer Science University of Sheffield Sheffield UK
2 shared publications
Latest Publications
- Open Access
- 7 February 2026
- Published by Elsevier in SSRN Electronic Journal
A Distributed Lightweight Edge Traffic Spatio-Temporal Forecasting Framework with Regional Autonomy and Global Feature Generation
While edge computing enables roadside traffic prediction, deploying complex deep models remains challenging due to high computational costs. Additionally, edge units face difficulties in improving prediction accuracy with limited access to cross-regional or global traffic information. To address these challenges, this paper proposes a distributed, lightweight edge traffic spatio-temporal prediction framework (DLiET) based on graph partitioning and knowledge distillation. The framework aims to achieve regional autonomy, load distribution, and global prediction. By decomposing the overall traffic graph into subgraphs with local models and a global model, DLiET enables efficient distributed computation and compensates for the loss of cross-regional information. Specifically, a static traffic graph partitioning process is introduced, alongside lightweight subgraph prediction models and a global prediction model based on a spatio-temporal aggregation module. Knowledge distillation is employed to maintain high prediction accuracy. The interaction between models is carefully designed to minimize communication costs. To address missing edge connections between subgraphs, a global feature generator based on truncated edge indexing is proposed, which collaborates with both the subgraph and global models. Experimental results across six traffic prediction datasets demonstrate that the DLiET model outperforms baseline models, with significantly fewer parameters and FLOPs than the compared models. The code for this paper is available at https://github.com/ouyangnann/DLiET.
- 7 February 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA Distributed Lightweight Edge Traffic Spatio-Temporal Forecasting Framework with Regional Autonomy and Global Feature Generation
While edge computing enables roadside traffic prediction, deploying complex deep models remains challenging due to high computational costs. Additionally, edge units face difficulties in improving prediction accuracy with limited access to cross-regional or global traffic information. To address these challenges, this paper proposes a distributed, lightweight edge traffic spatio-temporal prediction framework (DLiET) based on graph partitioning and knowledge distillation. The framework aims to achieve regional autonomy, load distribution, and global prediction. By decomposing the overall traffic graph into subgraphs with local models and a global model, DLiET enables efficient distributed computation and compensates for the loss of cross-regional information. Specifically, a static traffic graph partitioning process is introduced, alongside lightweight subgraph prediction models and a global prediction model based on a spatio-temporal aggregation module. Knowledge distillation is employed to maintain high prediction accuracy. The interaction between models is carefully designed to minimize communication costs. To address missing edge connections between subgraphs, a global feature generator based on truncated edge indexing is proposed, which collaborates with both the subgraph and global models. Experimental results across six traffic prediction datasets demonstrate that the DLiET model outperforms baseline models, with significantly fewer parameters and FLOPs than the compared models. The code for this paper is available at https://github.com/ouyangnann/DLiET.
- 15 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 22 (4) , 3376-3387
Saliency-Guided Transformer Attention With Pixel-Level Contrastive Learning for Weakly Supervised Defect Localization
Weakly supervised learning has emerged as a powerful paradigm for image segmentation, providing a practical solution to reduce the dependence on costly, pixel-level annotated datasets. By leveraging partial annotations, such as image-level tags, this approach significantly reduces the labeling burden while maintaining competitive performance. However, weakly supervised methods inherently face challenges, including overlapping activation regions and insufficient localization due to sparse and noisy signals from image-level tags, often leading to suboptimal segmentation performance. These issues are further exacerbated in industrial defect localization, where datasets present unique complexities, including low-contrast object boundaries, inconsistent shapes, interclass similarities, and substantial intraclass variations. To address these limitations, we introduce a novel saliency-guided transformer attention with contrastive learning framework. The proposed framework leverages Transformer attention to generate localization maps and enhances the learning of challenging foreground and background information through saliency-guided cues. In addition, a pixel-level contrastive learning module is employed to refine feature map representations by bringing positive pairs closer together and pushing negative pairs apart, effectively addressing challenges, such as overlapping activations and ambiguous boundaries. Through extensive experiments and ablation studies, the proposed method demonstrates superior performance compared to state-of-the-art approaches across three defect segmentation datasets. We also evaluated the generalization of our approach on the PASCAL VOC segmentation and MVTec anomaly detection datasets.
- 15 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 22 (4) , 3376-3387
Saliency-Guided Transformer Attention With Pixel-Level Contrastive Learning for Weakly Supervised Defect Localization
Weakly supervised learning has emerged as a powerful paradigm for image segmentation, providing a practical solution to reduce the dependence on costly, pixel-level annotated datasets. By leveraging partial annotations, such as image-level tags, this approach significantly reduces the labeling burden while maintaining competitive performance. However, weakly supervised methods inherently face challenges, including overlapping activation regions and insufficient localization due to sparse and noisy signals from image-level tags, often leading to suboptimal segmentation performance. These issues are further exacerbated in industrial defect localization, where datasets present unique complexities, including low-contrast object boundaries, inconsistent shapes, interclass similarities, and substantial intraclass variations. To address these limitations, we introduce a novel saliency-guided transformer attention with contrastive learning framework. The proposed framework leverages Transformer attention to generate localization maps and enhances the learning of challenging foreground and background information through saliency-guided cues. In addition, a pixel-level contrastive learning module is employed to refine feature map representations by bringing positive pairs closer together and pushing negative pairs apart, effectively addressing challenges, such as overlapping activations and ambiguous boundaries. Through extensive experiments and ablation studies, the proposed method demonstrates superior performance compared to state-of-the-art approaches across three defect segmentation datasets. We also evaluated the generalization of our approach on the PASCAL VOC segmentation and MVTec anomaly detection datasets.
- 16 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. 75 (3) , 3725-3739
DSTIFormer: Vehicle Trajectory Prediction Considering Dynamic Spatial-Temporal Interaction-Aware With Transformer Network
Current vehicle trajectory prediction methods often fail to capture the complex spatial and temporal interactions among surrounding vehicles in dynamic environments, leading to suboptimal prediction accuracy. In response to these challenges, we propose a novel Transformer-based model for vehicle trajectory prediction, denoted as the Dynamic Spatial-Temporal Interaction-aware Transformer Network (DSTIFormer). The three main parts of the DSTIFormer architecture are a spatial-temporal embedding module, a spatial-temporal dynamics interaction-aware module, and a trajectory encoderdecoder module. The spatial-temporal embedding module aims to encode vehicular trajectory interactions into latent representations that simultaneously preserve spatial dependencies and temporal features. The second module utilizes a temporal transformer network (TTN) along with a spatial graph transformer network (STN) to capture the dynamic spatial-temporal interactions between vehicles. To make trajectory predictions more accurate, the third module uses a sequence-to-sequence (Seq2Seq) architecture as the encoder and an attention-based decoder layer to effectively extract the long-range sequential dependencies in the process. The model's prediction accuracy with the root mean squared error (RMSE) metric on two publicly available datasets: NGSIM and HighD, and the results demonstrate that our proposed model outperforms several benchmark methods in long-term prediction.
- 16 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. 75 (3) , 3725-3739
DSTIFormer: Vehicle Trajectory Prediction Considering Dynamic Spatial-Temporal Interaction-Aware With Transformer Network
Current vehicle trajectory prediction methods often fail to capture the complex spatial and temporal interactions among surrounding vehicles in dynamic environments, leading to suboptimal prediction accuracy. In response to these challenges, we propose a novel Transformer-based model for vehicle trajectory prediction, denoted as the Dynamic Spatial-Temporal Interaction-aware Transformer Network (DSTIFormer). The three main parts of the DSTIFormer architecture are a spatial-temporal embedding module, a spatial-temporal dynamics interaction-aware module, and a trajectory encoderdecoder module. The spatial-temporal embedding module aims to encode vehicular trajectory interactions into latent representations that simultaneously preserve spatial dependencies and temporal features. The second module utilizes a temporal transformer network (TTN) along with a spatial graph transformer network (STN) to capture the dynamic spatial-temporal interactions between vehicles. To make trajectory predictions more accurate, the third module uses a sequence-to-sequence (Seq2Seq) architecture as the encoder and an attention-based decoder layer to effectively extract the long-range sequential dependencies in the process. The model's prediction accuracy with the root mean squared error (RMSE) metric on two publicly available datasets: NGSIM and HighD, and the results demonstrate that our proposed model outperforms several benchmark methods in long-term prediction.
- 12 August 2023
- Published by Elsevier in Computers in Industry
- Vol. 152
Uncertainty-aware and dynamically-mixed pseudo-labels for semi-supervised defect segmentation
- 12 August 2023
- Published by Elsevier in Computers in Industry
- Vol. 152
Uncertainty-aware and dynamically-mixed pseudo-labels for semi-supervised defect segmentation
- Open Access
- 13 January 2023
- Published by Association for Computing Machinery (ACM)
- p. 73-78
Semi-supervised Defect Segmentation with Uncertainty-aware Pseudo-labels from Multi-branch Network
Semi-supervised learning methods have recently gained considerable attention for training deep learning networks with limited labeled samples and additional large label-free samples. Consistency regularization and pseudo-labeling methods are among the most widely used semi-supervised learning methods. However, unreliable pseudo labels will largely limit the model’s performance when learning from unlabeled images. To alleviate this problem, we propose uncertainty-rectified pseudo labels generated from dynamically mixing predictions of multiple decoders with a shared encoder network for semi-supervised defect segmentation. We estimated the uncertainty as the prediction discrepancy between the average prediction and the output of each decoder head. The estimated uncertainty then guides the consistency training as well as the pseudo-label-based supervision. The proposed method achieved significant performance improvement over the fully supervised baseline and other state-of-the-art semi-supervised segmentation methods on similar labeled data proportions. We also performed an extensive ablation study to demonstrate that the proposed method performs well under various setups.
- 13 January 2023
- Published by Association for Computing Machinery (ACM)
- p. 73-78
Open AccessSemi-supervised Defect Segmentation with Uncertainty-aware Pseudo-labels from Multi-branch Network
Semi-supervised learning methods have recently gained considerable attention for training deep learning networks with limited labeled samples and additional large label-free samples. Consistency regularization and pseudo-labeling methods are among the most widely used semi-supervised learning methods. However, unreliable pseudo labels will largely limit the model’s performance when learning from unlabeled images. To alleviate this problem, we propose uncertainty-rectified pseudo labels generated from dynamically mixing predictions of multiple decoders with a shared encoder network for semi-supervised defect segmentation. We estimated the uncertainty as the prediction discrepancy between the average prediction and the output of each decoder head. The estimated uncertainty then guides the consistency training as well as the pseudo-label-based supervision. The proposed method achieved significant performance improvement over the fully supervised baseline and other state-of-the-art semi-supervised segmentation methods on similar labeled data proportions. We also performed an extensive ablation study to demonstrate that the proposed method performs well under various setups.
- 20 December 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 19 (9) , 9535-9545
Semisupervised Defect Segmentation With Pairwise Similarity Map Consistency and Ensemble-Based Cross Pseudolabels
Deep learning-based automatic defect segmentation is one of the hot research areas in computer vision application for the task of intelligent industrial inspection. Recently, several state-of-the-art models for image segmentation task have been proposed. However, their high performance is vastly dependent on the availability of large set of labeled data, which is one of the hindering factors in achieving full potential with deep learning methods in industrial inspection. In this study, we propose a novel method based on pairwise similarity map consistency with ensemble-based cross-pseudo labels for semi-supervised defect segmentation that uses limited labeled samples while exploiting additional label-free samples. The proposed approach uses three network branches that are regularized by pairwise similarity map consistency, and each of them is supervised by the pseudo labels generated by ensemble of predictions of the other two networks for the unlabeled samples. The proposed method achieved significant performance improvement over the baseline of learning only from the labeled images and the current state-of-the-art semi-supervised methods. We perform ablation studies and extensive experiments on various parameters and components to demonstrate that our method achieved state-of-the-art results on three different datasets.
- 20 December 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 19 (9) , 9535-9545
Semisupervised Defect Segmentation With Pairwise Similarity Map Consistency and Ensemble-Based Cross Pseudolabels
Deep learning-based automatic defect segmentation is one of the hot research areas in computer vision application for the task of intelligent industrial inspection. Recently, several state-of-the-art models for image segmentation task have been proposed. However, their high performance is vastly dependent on the availability of large set of labeled data, which is one of the hindering factors in achieving full potential with deep learning methods in industrial inspection. In this study, we propose a novel method based on pairwise similarity map consistency with ensemble-based cross-pseudo labels for semi-supervised defect segmentation that uses limited labeled samples while exploiting additional label-free samples. The proposed approach uses three network branches that are regularized by pairwise similarity map consistency, and each of them is supervised by the pseudo labels generated by ensemble of predictions of the other two networks for the unlabeled samples. The proposed method achieved significant performance improvement over the baseline of learning only from the labeled images and the current state-of-the-art semi-supervised methods. We perform ablation studies and extensive experiments on various parameters and components to demonstrate that our method achieved state-of-the-art results on three different datasets.