Hongjian Zhan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Hongjian Zhan has a total of 51 co-authors. Here are the top co-authors by number of shared publications:
- Yue Lu
East China Normal University
15 shared publications - Umapada Pal
Indian Statistical Institute
7 shared publications Yue Lu
5 shared publications- Bing Yin
State Key Laboratory of Cognitive Intelligence, Hefei, China and iFLYTEK Research, iFLYTEK, Hefei, China
4 shared publications - Dajian Zhong
Shanghai Maritime University
3 shared publications - Jiajun Wei
Shanghai Key Laboratory of Multidimensional Information Processing, East China Normal University
3 shared publications
Latest Publications
- Open Access
- 29 September 2026
- Published by Elsevier in SSRN Electronic Journal
GMS-RAG: Agentic Graph-Memory RAG with Modality-Stratified Retrieval
Retrieval-Augmented Generation (RAG) has been widely utilized to integrate external knowledge into Large Language Models (LLMs) for enhancement onmultiple tasks. Recently, graph-guided RAG approaches have demonstrated stronger support for more accurate, context-aware responses. However, traditional RAG methods rely on static pipelines, failing to adaptively balance heterogeneous modalities on complex tasks. To address this issue, we propose GMS-RAG, a novel agentic RAG framework. GMS-RAG operates via a bi-stage autonomous agent system: a Graph-Memory Construction Agent and a Modality-Stratified Retrieval Agent. The construction stage first integrates visual contents along with textual ones into a unified graph memory, then a balanced retrieval mechanism is employed to aggregate evidence for answer generation. Experiments on four benchmarks indicate that GMS-RAG effectively enhances the multi-modal RAG capability of backbone models. Performance measured by F1 score shows that our framework outperforms all baselines with improved quality of answer generation and generalizability over modalities. The code will be available soon.
- 29 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessGMS-RAG: Agentic Graph-Memory RAG with Modality-Stratified Retrieval
Retrieval-Augmented Generation (RAG) has been widely utilized to integrate external knowledge into Large Language Models (LLMs) for enhancement onmultiple tasks. Recently, graph-guided RAG approaches have demonstrated stronger support for more accurate, context-aware responses. However, traditional RAG methods rely on static pipelines, failing to adaptively balance heterogeneous modalities on complex tasks. To address this issue, we propose GMS-RAG, a novel agentic RAG framework. GMS-RAG operates via a bi-stage autonomous agent system: a Graph-Memory Construction Agent and a Modality-Stratified Retrieval Agent. The construction stage first integrates visual contents along with textual ones into a unified graph memory, then a balanced retrieval mechanism is employed to aggregate evidence for answer generation. Experiments on four benchmarks indicate that GMS-RAG effectively enhances the multi-modal RAG capability of backbone models. Performance measured by F1 score shows that our framework outperforms all baselines with improved quality of answer generation and generalizability over modalities. The code will be available soon.
- 1 November 2025
- Published by Springer Nature
- p. 247-259
A Local Perceptual Approach for Few-Shot Text Effect Transfer
- 1 November 2025
- Published by Springer Nature
- p. 247-259
A Local Perceptual Approach for Few-Shot Text Effect Transfer
- 12 September 2025
- Published by Elsevier in Engineering Applications of Artificial Intelligence
- Vol. 161
Length-aware center loss for sequence to sequence Thai scene text recognition
- 12 September 2025
- Published by Elsevier in Engineering Applications of Artificial Intelligence
- Vol. 161
Length-aware center loss for sequence to sequence Thai scene text recognition
- 7 December 2024
- Published by Springer Nature
- p. 81-92
FaRE: A Feature-Aware Radical Encoding Strategy for Zero-Shot Chinese Character Recognition
- 7 December 2024
- Published by Springer Nature
- p. 81-92
FaRE: A Feature-Aware Radical Encoding Strategy for Zero-Shot Chinese Character Recognition
- 4 December 2024
- Published by Springer Nature
- p. 275-290
LK-Net: Efficient Large Kernel ConvNet for Document Enhancement
- 4 December 2024
- Published by Springer Nature
- p. 275-290
LK-Net: Efficient Large Kernel ConvNet for Document Enhancement
- 4 December 2024
- Published by Springer Nature
- p. 360-374
RSTAN: Residual Spatio-Temporal Attention Network for End-to-End Human Fall Detection
- 4 December 2024
- Published by Springer Nature
- p. 360-374
RSTAN: Residual Spatio-Temporal Attention Network for End-to-End Human Fall Detection
- 30 November 2024
- Published by Springer Nature
- p. 94-109
Learning to Detect Lithography Defects in SEM Images
- 30 November 2024
- Published by Springer Nature
- p. 94-109
Learning to Detect Lithography Defects in SEM Images
- Open Access
- 28 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 6202-6211
Free Lunch: Frame-level Contrastive Learning with Text Perceiver for Robust Scene Text Recognition in Lightweight Models
Lightweight models play an important role in real-life applications, especially in the recent mobile device era. However, due to limited network scale and low-quality images, the performance of lightweight models on Scene Text Recognition (STR) tasks is still much to be improved. Recently, contrastive learning has shown its power in many areas, with promising performances without additional computational cost. Based on these observations, we propose a new efficient and effective frame-level contrastive learning (FLCL) framework for lightweight STR models. The FLCL framework consists of a backbone to extract basic features, a Text Perceiver Module (TPM) to focus on text-relevant representations, and a FLCL loss to update the network. The backbone can be any feature extraction architecture. The TPM is an innovative Mamba-based structure that is designed to suppress features irrelevant to the text content from the backbone. Unlike existing word-level contrastive learning, we look into the nature of the STR task and propose the frame-level contrastive learning loss, which can work well with the famous Connectionist Temporal Classification loss. We conduct experiments on six well-known STR benchmarks as well as a new low-quality dataset. Compared to vanilla contrastive learning and other non-parameter methods, the FLCL framework significantly outperforms others on all datasets, especially the low-quality dataset. In addition, character feature visualization demonstrates that the proposed method can yield more discriminative character features for visually similar characters, which also substantiates the efficacy of the proposed methods. Codes and the low-quality dataset will be available soon.
- 28 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 6202-6211
Open AccessFree Lunch: Frame-level Contrastive Learning with Text Perceiver for Robust Scene Text Recognition in Lightweight Models
Lightweight models play an important role in real-life applications, especially in the recent mobile device era. However, due to limited network scale and low-quality images, the performance of lightweight models on Scene Text Recognition (STR) tasks is still much to be improved. Recently, contrastive learning has shown its power in many areas, with promising performances without additional computational cost. Based on these observations, we propose a new efficient and effective frame-level contrastive learning (FLCL) framework for lightweight STR models. The FLCL framework consists of a backbone to extract basic features, a Text Perceiver Module (TPM) to focus on text-relevant representations, and a FLCL loss to update the network. The backbone can be any feature extraction architecture. The TPM is an innovative Mamba-based structure that is designed to suppress features irrelevant to the text content from the backbone. Unlike existing word-level contrastive learning, we look into the nature of the STR task and propose the frame-level contrastive learning loss, which can work well with the famous Connectionist Temporal Classification loss. We conduct experiments on six well-known STR benchmarks as well as a new low-quality dataset. Compared to vanilla contrastive learning and other non-parameter methods, the FLCL framework significantly outperforms others on all datasets, especially the low-quality dataset. In addition, character feature visualization demonstrates that the proposed method can yield more discriminative character features for visually similar characters, which also substantiates the efficacy of the proposed methods. Codes and the low-quality dataset will be available soon.
- 18 October 2024
- Published by Elsevier in Expert Systems with Applications
- Vol. 261
VGTS: Visually Guided Text Spotting for novel categories in historical manuscripts
- 18 October 2024
- Published by Elsevier in Expert Systems with Applications
- Vol. 261
VGTS: Visually Guided Text Spotting for novel categories in historical manuscripts
- 16 May 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Audio, Speech, and Language Processing
- Vol. 32 (23299290) , 2814-2827
DS-TDNN: Dual-Stream Time-Delay Neural Network With Global-Aware Filter for Speaker Verification
Conventional time-delay neural networks (TDNNs) struggle to handle long-range context, their ability to represent speaker information is therefore limited for long utterances. Existing solutions either depend on increasing model complexity or try to strike a balance between local features and global context to address this issue. To effectively leverage the long-term dependencies of audio signals and constrain model complexity, we introduce a novel module called Global-aware Filter layer (GF layer) in this work, which employs a set of learnable transform-domain filters between a 1D discrete Fourier transform and its inverse transform to capture global context. Additionally, we develop a dynamic filtering strategy and a sparse regularization method to enhance the performance of the GF layer and prevent overfitting. Based on the GF layer, we present a dual-stream TDNN architecture called DS-TDNN for automatic speaker verification (ASV), which utilizes two unique branches to extract both local and global features in parallel and employs an efficient strategy to fuse different-scale information. Experiments on the Voxceleb and SITW databases demonstrate that the DS-TDNN achieves a relative improvement of 10% together with a relative decline of 20% in computational cost over the ECAPA-TDNN in the speaker verification task. This improvement becomes more evident as the utterance's duration grows. Furthermore, the DS-TDNN also beats popular deep residual models and attention-based systems on utterances of arbitrary length.
- 16 May 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Audio, Speech, and Language Processing
- Vol. 32 (23299290) , 2814-2827
DS-TDNN: Dual-Stream Time-Delay Neural Network With Global-Aware Filter for Speaker Verification
Conventional time-delay neural networks (TDNNs) struggle to handle long-range context, their ability to represent speaker information is therefore limited for long utterances. Existing solutions either depend on increasing model complexity or try to strike a balance between local features and global context to address this issue. To effectively leverage the long-term dependencies of audio signals and constrain model complexity, we introduce a novel module called Global-aware Filter layer (GF layer) in this work, which employs a set of learnable transform-domain filters between a 1D discrete Fourier transform and its inverse transform to capture global context. Additionally, we develop a dynamic filtering strategy and a sparse regularization method to enhance the performance of the GF layer and prevent overfitting. Based on the GF layer, we present a dual-stream TDNN architecture called DS-TDNN for automatic speaker verification (ASV), which utilizes two unique branches to extract both local and global features in parallel and employs an efficient strategy to fuse different-scale information. Experiments on the Voxceleb and SITW databases demonstrate that the DS-TDNN achieves a relative improvement of 10% together with a relative decline of 20% in computational cost over the ECAPA-TDNN in the speaker verification task. This improvement becomes more evident as the utterance's duration grows. Furthermore, the DS-TDNN also beats popular deep residual models and attention-based systems on utterances of arbitrary length.