Sulan Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Sulan Zhang has a total of 49 co-authors. Here are the top co-authors by number of shared publications:
- Jifu Zhang
Taiyuan University of Science and Technology
10 shared publications - Lihua Hu
School of Computer Science and Technology, Taiyuan University of Science and Technology, Taiyuan, China
7 shared publications Jifu Zhang
School of Computer Science and Technology, Taiyuan University of Science and Technology, Taiyuan, China
5 shared publicationsLihua Hu
5 shared publications- Ping Guo
Beijing Normal University
5 shared publications Huiyuan Zhou
3 shared publications
Latest Publications
- Open Access
- 13 November 2026
- Published by MDPI AG in Mathematics
- Vol. 14 (22) , 4172
Dynamic Adaptive Random Projection with Stacking Pseudoinverse Learning for Image Classification
Random projection is a vital dimensionality-reduction technique for high-dimensional data, whose performance heavily relies on the distribution design of its projection matrix. Conventional approaches generate projection matrices with predefined fixed distributions and thus fail to adapt to the complex distributions inherent in image data. This deficiency frequently causes feature loss or computational redundancy, which degrades overall model performance. To tackle this problem, this paper proposes a dynamic adaptive random projection with stacking pseudoinverse learning (DARPSP) for image classification. The core of the proposed method lies in building a fully dynamic adaptive random-projection mechanism. First, a data-feature perception module dynamically captures the underlying data distribution. Second, a parameterized dynamic distribution-control network predicts distribution parameters for the projection matrix to realize adaptive adjustment. A fusion strategy based on feature-importance evaluation is then adopted to refine the combination of multi-distribution projection outputs. Finally, the above components are integrated into a single-hidden-layer stacking pseudoinverse learning architecture to further boost model performance. The experimental results on the MNIST, COIL-20, and Salinas public datasets show that the proposed method outperforms the comparison methods on accuracy, precision, recall, and F1.
- 13 November 2026
- Published by MDPI AG in Mathematics
- Vol. 14 (22) , 4172
Open AccessDynamic Adaptive Random Projection with Stacking Pseudoinverse Learning for Image Classification
Random projection is a vital dimensionality-reduction technique for high-dimensional data, whose performance heavily relies on the distribution design of its projection matrix. Conventional approaches generate projection matrices with predefined fixed distributions and thus fail to adapt to the complex distributions inherent in image data. This deficiency frequently causes feature loss or computational redundancy, which degrades overall model performance. To tackle this problem, this paper proposes a dynamic adaptive random projection with stacking pseudoinverse learning (DARPSP) for image classification. The core of the proposed method lies in building a fully dynamic adaptive random-projection mechanism. First, a data-feature perception module dynamically captures the underlying data distribution. Second, a parameterized dynamic distribution-control network predicts distribution parameters for the projection matrix to realize adaptive adjustment. A fusion strategy based on feature-importance evaluation is then adopted to refine the combination of multi-distribution projection outputs. Finally, the above components are integrated into a single-hidden-layer stacking pseudoinverse learning architecture to further boost model performance. The experimental results on the MNIST, COIL-20, and Salinas public datasets show that the proposed method outperforms the comparison methods on accuracy, precision, recall, and F1.
- 12 September 2026
- Published by Elsevier in Computers & Graphics
- Vol. 140
C-PGS: Causal prior-guided style-aware semi-supervised semantic segmentation for ancient Chinese architecture
- 12 September 2026
- Published by Elsevier in Computers & Graphics
- Vol. 140
C-PGS: Causal prior-guided style-aware semi-supervised semantic segmentation for ancient Chinese architecture
- Open Access
- 26 February 2026
- Published by MDPI AG in Electronics
- Vol. 15 (5) , 964
RAFFNet: Restricted Attention Feature Fusion Network for Self-Supervised Image Representation Learning
Learning image representations with deep self-supervised models is an important task in computer vision, which aims to establish beneficial and general representations from unlabeled images. However, existing efforts train models mainly on high-level features, neglecting lower-level features and their global spatial information, thus limiting the discriminative power of the learned representations. In this work, we propose a representational learning model based on restricted attention feature fusion network (RAFFNet) to improve the quality and generalization of the learned image representations. Specifically, to fully exploit the features in the deep network, we use a self-supervised model on multi-level features to learn more general representations. Meanwhile, a new feature fusion strategy with a dual attention mechanism of channel and space is used for multi-level features, enabling the model training to obtain more important and comprehensive feature information. Furthermore, in order to better extract global spatial information, we devise a simple but effective attentional weighted mask, which restricts the weight of spatial attention and prevents the model from focusing only on local features with high attention weights. Experiments on four public classification datasets, CIFAR-10, CIFAR-100, Tiny ImageNet and ImageNet-1%, and two object detection datasets, PASCAL VOC and COCO, demonstrate that the proposed RAFFNet has better representation performance and generalization ability than most state-of-the-art image representation learning algorithms.
- 26 February 2026
- Published by MDPI AG in Electronics
- Vol. 15 (5) , 964
Open AccessRAFFNet: Restricted Attention Feature Fusion Network for Self-Supervised Image Representation Learning
Learning image representations with deep self-supervised models is an important task in computer vision, which aims to establish beneficial and general representations from unlabeled images. However, existing efforts train models mainly on high-level features, neglecting lower-level features and their global spatial information, thus limiting the discriminative power of the learned representations. In this work, we propose a representational learning model based on restricted attention feature fusion network (RAFFNet) to improve the quality and generalization of the learned image representations. Specifically, to fully exploit the features in the deep network, we use a self-supervised model on multi-level features to learn more general representations. Meanwhile, a new feature fusion strategy with a dual attention mechanism of channel and space is used for multi-level features, enabling the model training to obtain more important and comprehensive feature information. Furthermore, in order to better extract global spatial information, we devise a simple but effective attentional weighted mask, which restricts the weight of spatial attention and prevents the model from focusing only on local features with high attention weights. Experiments on four public classification datasets, CIFAR-10, CIFAR-100, Tiny ImageNet and ImageNet-1%, and two object detection datasets, PASCAL VOC and COCO, demonstrate that the proposed RAFFNet has better representation performance and generalization ability than most state-of-the-art image representation learning algorithms.
- 28 January 2026
- Published by Elsevier in Knowledge-Based Systems
- Vol. 337
Multimodal interpretable image recognition network via language-guided global-local collaboratively alignment
- 28 January 2026
- Published by Elsevier in Knowledge-Based Systems
- Vol. 337
Multimodal interpretable image recognition network via language-guided global-local collaboratively alignment
- 17 January 2026
- Published by Springer Nature in Pattern Analysis and Applications
- Vol. 29 (1) , 28
CFG-NeRF: coordinate-feature-gate collaborative optimization for sparse-view ancient architecture reconstruction
- 17 January 2026
- Published by Springer Nature in Pattern Analysis and Applications
- Vol. 29 (1) , 28
CFG-NeRF: coordinate-feature-gate collaborative optimization for sparse-view ancient architecture reconstruction
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 28, 12255-12266
Hyperbolic Concept Bottleneck Model for Interpretable Fine-Grained Image Classification
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 28, 12255-12266
Hyperbolic Concept Bottleneck Model for Interpretable Fine-Grained Image Classification
- Open Access
- 15 February 2025
- Published by Springer Nature in npj Heritage Science
- Vol. 13 (1) , 1-13
A coordinate-to-semantic attention network for multi-label ancient Chinese architecture image classification
- 15 February 2025
- Published by Springer Nature in npj Heritage Science
- Vol. 13 (1) , 1-13
Open AccessA coordinate-to-semantic attention network for multi-label ancient Chinese architecture image classification
- Open Access
- 1 January 2025
- Published by Institution of Engineering and Technology (IET) in IET Image Processing
- Vol. 19 (1)
Two‐Stage Feature Selection for Fine‐Grained Image Recognition Via Partial Order Analysis and Heterogeneity Evaluation
The core challenge of fine‐grained image recognition (FGIR) tasks is distinguishing highly similar subclasses within the same base category. Most CNN‐based deep learning methods typically focus on extracting information from local regions while overlook the inherent structure between subclasses and the complex relationships between features. This paper presents a two‐stage feature selection method based on partial order analysis (POA) and heterogeneity evaluation (HE) for FGIR tasks, guiding the model to focus on distinctive features while reducing uncertainty caused by interfering information. Specifically, in the POA stage, clustering first groups similar subcategories into a medium‐granularity category. Formal concept analysis then models their hierarchical partial order, identifying “shared features” among subcategories and “exclusive features” unique to each. This structured representation highlights key contrastive cues. In the HE stage, a novel heterogeneity index is introduced to measure the fluctuation of low‐level features within each fine‐grained category. This index guides the model to suppress pseudo‐discriminative features with high heterogeneity, mitigating the impact of noisy and unstable information on decision‐making. We perform comprehensive experiments on three commonly used benchmark datasets (CUB‐200‐2011, Stanford Cars, and FGVC‐Aircraft). Experimental results show that the proposed method outperforms classic FGIC methods, validating the effectiveness of our approach.
- 1 January 2025
- Published by Institution of Engineering and Technology (IET) in IET Image Processing
- Vol. 19 (1)
Open AccessTwo‐Stage Feature Selection for Fine‐Grained Image Recognition Via Partial Order Analysis and Heterogeneity Evaluation
The core challenge of fine‐grained image recognition (FGIR) tasks is distinguishing highly similar subclasses within the same base category. Most CNN‐based deep learning methods typically focus on extracting information from local regions while overlook the inherent structure between subclasses and the complex relationships between features. This paper presents a two‐stage feature selection method based on partial order analysis (POA) and heterogeneity evaluation (HE) for FGIR tasks, guiding the model to focus on distinctive features while reducing uncertainty caused by interfering information. Specifically, in the POA stage, clustering first groups similar subcategories into a medium‐granularity category. Formal concept analysis then models their hierarchical partial order, identifying “shared features” among subcategories and “exclusive features” unique to each. This structured representation highlights key contrastive cues. In the HE stage, a novel heterogeneity index is introduced to measure the fluctuation of low‐level features within each fine‐grained category. This index guides the model to suppress pseudo‐discriminative features with high heterogeneity, mitigating the impact of noisy and unstable information on decision‐making. We perform comprehensive experiments on three commonly used benchmark datasets (CUB‐200‐2011, Stanford Cars, and FGVC‐Aircraft). Experimental results show that the proposed method outperforms classic FGIC methods, validating the effectiveness of our approach.
- 7 December 2024
- Published by Springer Nature in Journal of Real-Time Image Processing
- Vol. 22 (1) , 1-14
Towards real-time accurate dense pedestrian detection via large-kernel perception module and multi-level feature fusion
- 7 December 2024
- Published by Springer Nature in Journal of Real-Time Image Processing
- Vol. 22 (1) , 1-14
Towards real-time accurate dense pedestrian detection via large-kernel perception module and multi-level feature fusion
- Open Access
- 11 May 2024
- Published by Springer Nature in Complex & Intelligent Systems
- Vol. 10 (4) , 5569-5583
EMC+GD_C: circle-based enhanced motion consistency and guided diffusion feature matching for 3D reconstruction
Robust matching, especially the number, precision and distribution of feature point matching, directly affects the effect of 3D reconstruction. However, the existing methods rarely consider these three aspects comprehensively to improve the quality of feature matching, which in turn affects the effect of 3D reconstruction. Therefore, to effectively improve the quality of 3D reconstruction, we propose a circle-based enhanced motion consistency and guided diffusion feature matching algorithm for 3D reconstruction named EMC+GD_C. Firstly, a circle-based neighborhood division method is proposed, which increases the number of initial matching points. Secondly, to improve the precision of feature point matching, on the one hand, we put forward the idea of enhancing motion consistency, reducing the mismatch of high similarity feature points by enhancing the judgment conditions of true and false matching points; on the other hand, we combine the RANSAC optimization method to filter out the outliers and further improve the precision of feature point matching. Finally, a novel guided diffusion idea combining guided matching and motion consistency is proposed, which expands the distribution range of feature point matching and improves the stability of 3D models. Experiments on 8 sets of 908 pairs of images in the public 3D reconstruction datasets demonstrate that our method can achieve better matching performance and show stronger stability in 3D reconstruction. Specifically, EMC+GD_C achieves an average improvement of 24.07% compared to SIFT-based ratio test, 9.18% to GMS and 1.94% to EMC+GD_G in feature matching precision.
- 11 May 2024
- Published by Springer Nature in Complex & Intelligent Systems
- Vol. 10 (4) , 5569-5583
Open AccessEMC+GD_C: circle-based enhanced motion consistency and guided diffusion feature matching for 3D reconstruction
Robust matching, especially the number, precision and distribution of feature point matching, directly affects the effect of 3D reconstruction. However, the existing methods rarely consider these three aspects comprehensively to improve the quality of feature matching, which in turn affects the effect of 3D reconstruction. Therefore, to effectively improve the quality of 3D reconstruction, we propose a circle-based enhanced motion consistency and guided diffusion feature matching algorithm for 3D reconstruction named EMC+GD_C. Firstly, a circle-based neighborhood division method is proposed, which increases the number of initial matching points. Secondly, to improve the precision of feature point matching, on the one hand, we put forward the idea of enhancing motion consistency, reducing the mismatch of high similarity feature points by enhancing the judgment conditions of true and false matching points; on the other hand, we combine the RANSAC optimization method to filter out the outliers and further improve the precision of feature point matching. Finally, a novel guided diffusion idea combining guided matching and motion consistency is proposed, which expands the distribution range of feature point matching and improves the stability of 3D models. Experiments on 8 sets of 908 pairs of images in the public 3D reconstruction datasets demonstrate that our method can achieve better matching performance and show stronger stability in 3D reconstruction. Specifically, EMC+GD_C achieves an average improvement of 24.07% compared to SIFT-based ratio test, 9.18% to GMS and 1.94% to EMC+GD_G in feature matching precision.