Cencen Liu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Cencen Liu has a total of 31 co-authors. Here are the top co-authors by number of shared publications:
- Dongyang Zhang
University of Electronic Science and Technology of China, Chengdu, Sichuan, China
6 shared publications - Guoming Lu
University of Electronic Science and Technology of China, Chengdu, China
4 shared publications - Wen Yin
University of Electronic Science and Technology of China
4 shared publications - Dezhang Zheng
University of Electronic Science and Technology of China
2 shared publications Jielei Wang
2 shared publications- Jielei Wang
University of Electronic Science and Technology of China
2 shared publications
Latest Publications
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 6128-6137
What to Remove, What to Preserve: Dual-Ambiguity Rectification for All-in-One Image Restoration
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 6128-6137
Open AccessWhat to Remove, What to Preserve: Dual-Ambiguity Rectification for All-in-One Image Restoration
- Open Access
- 15 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 440-448
DTSR: High-Frequency Prior-Based Dynamic Texture Synthesis for Real-World Image Super-Resolution
- 15 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 440-448
Open AccessDTSR: High-Frequency Prior-Based Dynamic Texture Synthesis for Real-World Image Super-Resolution
- Open Access
- 6 April 2026
- Published by Elsevier in SSRN Electronic Journal
HMT-SR: Hybrid Mamba-Transformer with Frequency Awareness for Single Image Super-Resolution
While Transformer and Mamba architectures have significantly advanced Single Image Super-Resolution (SISR), they suffer from distinct spectral mismatches: Transformers allocate disproportionate computational resources to low-frequency redundancies, whereas Mamba's recursive mechanism inherently acts as a low-pass filter, severely attenuating essential high-frequency details. Existing hybrid approaches naively stack these blocks within a unified feature space, leaving this spectral entanglement unresolved. To address this, we propose HMT-SR, a novel frequency-aware framework that explicitly decouples spectral components to optimally align with their respective architectural inductive biases. Following an initial Frequency-Decoupling Module (FDM), HMT-SR routes features into specialized streams: (i) a Structure-Preserving Mamba (SPM) to construct the low-frequency global structural foundation; (ii) an Adaptive High-Frequency Enhancement (AHFE) module to efficiently recover fine textures, directly compensating for Mamba's high-frequency loss; and (iii) a Detail-Guided Transformer (DGT) serving as a high-fidelity fusion engine. Extensive experiments demonstrate that HMT-SR achieves an exceptional quality-efficiency trade-off. Specifically, compared to the state-of-the-art MambaIRv2-light, HMT-SR requires fewer parameters while achieving an average PSNR gain of 0.16 dB on the Manga109 benchmark.
- 6 April 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessHMT-SR: Hybrid Mamba-Transformer with Frequency Awareness for Single Image Super-Resolution
While Transformer and Mamba architectures have significantly advanced Single Image Super-Resolution (SISR), they suffer from distinct spectral mismatches: Transformers allocate disproportionate computational resources to low-frequency redundancies, whereas Mamba's recursive mechanism inherently acts as a low-pass filter, severely attenuating essential high-frequency details. Existing hybrid approaches naively stack these blocks within a unified feature space, leaving this spectral entanglement unresolved. To address this, we propose HMT-SR, a novel frequency-aware framework that explicitly decouples spectral components to optimally align with their respective architectural inductive biases. Following an initial Frequency-Decoupling Module (FDM), HMT-SR routes features into specialized streams: (i) a Structure-Preserving Mamba (SPM) to construct the low-frequency global structural foundation; (ii) an Adaptive High-Frequency Enhancement (AHFE) module to efficiently recover fine textures, directly compensating for Mamba's high-frequency loss; and (iii) a Detail-Guided Transformer (DGT) serving as a high-fidelity fusion engine. Extensive experiments demonstrate that HMT-SR achieves an exceptional quality-efficiency trade-off. Specifically, compared to the state-of-the-art MambaIRv2-light, HMT-SR requires fewer parameters while achieving an average PSNR gain of 0.16 dB on the Manga109 benchmark.
- 28 January 2026
- Published by Elsevier in Knowledge-Based Systems
- Vol. 337
Rethinking attention cues: Multi-Factor guided token pruning for efficient vision-language understanding
- 28 January 2026
- Published by Elsevier in Knowledge-Based Systems
- Vol. 337
Rethinking attention cues: Multi-Factor guided token pruning for efficient vision-language understanding
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Imaging
- Vol. 12 (25730436) , 787-799
HSFM: Hybrid Spatial Frequency Mamba for Efficient Image Restoration
Natural images often suffer from various degradations due to adverse weather conditions and imaging environments. Consequently, image restoration has developed as a crucial solution and continues to attract attention. Although recent transformer-based architectures have demonstrated remarkable success in image restoration, their substantial computational demands create considerable obstacles for real-world deployment. Vision Mamba demonstrates exceptional long-range dependency modeling capabilities with linear complexity, showcasing its potential to replace Transformers. However, single spatial domain information from Mamba tends to be usually insufficient, constraining the model's ability to handle complex real-world degradation scenarios. Furthermore, current methods commonly use MLP layers or incorporate convolution operations for feature propagation in the feed forward network, underutilizing the captured fine-grained domain characteristics. Inspired by the above insights, we propose an efficient network for image restoration, namely Hybrid Spatial Frequency Mamba Network (HSFM). Specifically, our network features two key designs: 1) We propose a Spatial Fourier Mixer (SFM) that integrates spatial features from a selective multi-scale interaction branch and frequency features derived through an adaptive context integration branch; 2) We design a Hybrid Feature Aggregation Module (HFAM) to enhance detail preservation to obtain better image restoration performance. Extensive experiments on three representative image restoration tasks, including image super-resolution, image denoising and image dehazing demonstrate that our network achieves superior performance in restoration quality while maintaining low complexity and memory consumption. The code is available at https://github.com/Qiuyu2024/HSFM.
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Imaging
- Vol. 12 (25730436) , 787-799
HSFM: Hybrid Spatial Frequency Mamba for Efficient Image Restoration
Natural images often suffer from various degradations due to adverse weather conditions and imaging environments. Consequently, image restoration has developed as a crucial solution and continues to attract attention. Although recent transformer-based architectures have demonstrated remarkable success in image restoration, their substantial computational demands create considerable obstacles for real-world deployment. Vision Mamba demonstrates exceptional long-range dependency modeling capabilities with linear complexity, showcasing its potential to replace Transformers. However, single spatial domain information from Mamba tends to be usually insufficient, constraining the model's ability to handle complex real-world degradation scenarios. Furthermore, current methods commonly use MLP layers or incorporate convolution operations for feature propagation in the feed forward network, underutilizing the captured fine-grained domain characteristics. Inspired by the above insights, we propose an efficient network for image restoration, namely Hybrid Spatial Frequency Mamba Network (HSFM). Specifically, our network features two key designs: 1) We propose a Spatial Fourier Mixer (SFM) that integrates spatial features from a selective multi-scale interaction branch and frequency features derived through an adaptive context integration branch; 2) We design a Hybrid Feature Aggregation Module (HFAM) to enhance detail preservation to obtain better image restoration performance. Extensive experiments on three representative image restoration tasks, including image super-resolution, image denoising and image dehazing demonstrate that our network achieves superior performance in restoration quality while maintaining low complexity and memory consumption. The code is available at https://github.com/Qiuyu2024/HSFM.
- Open Access
- 12 September 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Asian and Low-Resource Language Information Processing
- Vol. 24 (10) , 1-17
DDM4TST: Diffusion Model for Fine-grained Text Style Transfer by Disentangled Representation
Fine-grained Text Style Transfer (FTST) aims to make targeted and precise modifications to specific stylistic components of a sentence. Existing methods typically attempt to disentangle style and content representations for FTST. However, style and content are inherently abstract, making them difficult to formalize and manipulate independently. To address this, we propose a novel framework, Disentanglement Diffusion Model for Text Style Transfer (DDM4TST) , which reformulates style transformation as either semantic or syntactic transformation. By learning disentangled representations, the model enables accurate and fine-grained style control. Specifically, we construct a feature parsing module to effectively separate semantic and syntactic representations. We then incorporate a diffusion model conditioned on these disentangled representations, allowing for fine-grained control of stylistic attributes while preserving the core content of the original sentence. This integration into the denoising process enhances the controllability and precision of the style transformation. Extensive experiments on the benchmark StylePTB dataset demonstrate that our model consistently outperforms widely adopted baselines. The results validate the effectiveness of our approach in achieving high-quality style transformation while maintaining content fidelity.
- 12 September 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Asian and Low-Resource Language Information Processing
- Vol. 24 (10) , 1-17
Open AccessDDM4TST: Diffusion Model for Fine-grained Text Style Transfer by Disentangled Representation
Fine-grained Text Style Transfer (FTST) aims to make targeted and precise modifications to specific stylistic components of a sentence. Existing methods typically attempt to disentangle style and content representations for FTST. However, style and content are inherently abstract, making them difficult to formalize and manipulate independently. To address this, we propose a novel framework, Disentanglement Diffusion Model for Text Style Transfer (DDM4TST) , which reformulates style transformation as either semantic or syntactic transformation. By learning disentangled representations, the model enables accurate and fine-grained style control. Specifically, we construct a feature parsing module to effectively separate semantic and syntactic representations. We then incorporate a diffusion model conditioned on these disentangled representations, allowing for fine-grained control of stylistic attributes while preserving the core content of the original sentence. This integration into the denoising process enhances the controllability and precision of the style transformation. Extensive experiments on the benchmark StylePTB dataset demonstrate that our model consistently outperforms widely adopted baselines. The results validate the effectiveness of our approach in achieving high-quality style transformation while maintaining content fidelity.
- Top Cited Papers
- 22 January 2025
- Published by Elsevier in Neurocomputing
- Vol. 624
SRMamba-T: Exploring the hybrid Mamba-Transformer network for Single Image Super-Resolution
- 22 January 2025
- Published by Elsevier in Neurocomputing
- Vol. 624
Top Cited PapersSRMamba-T: Exploring the hybrid Mamba-Transformer network for Single Image Super-Resolution
- Open Access
- 30 May 2024
- Published by Association for Computing Machinery (ACM)
- p. 1079-1083
Knowledge Distillation for Single Image Super-Resolution via Contrastive Learning
In recent years, thanks to the vigorous development of deep learning, single image super-resolution has advanced greatly. Most super-resolution (SR) methods use convolution layers to construct the network, which achieves superior results over the traditional methods based on manual features. However, most methods based on convolutional neural networks (CNN) blindly deepen the depth of the network leading to a large number of model parameters, which inevitably brings huge computing overhead and memory consumption, and greatly limits the application in resource-limited devices. In order to alleviate this problem, a knowledge distillation framework based on contrastive learning is proposed to compress and accelerate the SR model with enormous parameters. The student network is directly constructed by reducing the number of layers of the teacher network. In particular, the proposed method distills the statistical information of the intermediate feature maps from the teacher network to train the lightweight student network. In addition, through explicit knowledge transfer, a novel contrastive loss is introduced to improve the reconstruction performance of the student network. Experiments show that the proposed contrastive distillation framework can effectively compress the model scale with an acceptable loss of performance.
- 30 May 2024
- Published by Association for Computing Machinery (ACM)
- p. 1079-1083
Open AccessKnowledge Distillation for Single Image Super-Resolution via Contrastive Learning
In recent years, thanks to the vigorous development of deep learning, single image super-resolution has advanced greatly. Most super-resolution (SR) methods use convolution layers to construct the network, which achieves superior results over the traditional methods based on manual features. However, most methods based on convolutional neural networks (CNN) blindly deepen the depth of the network leading to a large number of model parameters, which inevitably brings huge computing overhead and memory consumption, and greatly limits the application in resource-limited devices. In order to alleviate this problem, a knowledge distillation framework based on contrastive learning is proposed to compress and accelerate the SR model with enormous parameters. The student network is directly constructed by reducing the number of layers of the teacher network. In particular, the proposed method distills the statistical information of the intermediate feature maps from the teacher network to train the lightweight student network. In addition, through explicit knowledge transfer, a novel contrastive loss is introduced to improve the reconstruction performance of the student network. Experiments show that the proposed contrastive distillation framework can effectively compress the model scale with an acceptable loss of performance.
- 22 September 2023
- Published by Springer Nature
- p. 505-517
Prompt-Oriented Fine-Tuning Dual Bert for Aspect-Based Sentiment Analysis
- 22 September 2023
- Published by Springer Nature
- p. 505-517
Prompt-Oriented Fine-Tuning Dual Bert for Aspect-Based Sentiment Analysis
- Open Access
- 25 August 2023
- Published by Association for Computing Machinery (ACM)
- p. 135-140
Structure-aware Table-to-Text Generation with Prefix-tuning
- 25 August 2023
- Published by Association for Computing Machinery (ACM)
- p. 135-140
Open AccessStructure-aware Table-to-Text Generation with Prefix-tuning