Tammam Tillo
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Tammam Tillo has a total of 15 co-authors. Here are the top co-authors by number of shared publications:
- Jimin Xiao
Department of Intelligent Science, Xian Jiaotong-Liverpool University, Suzhou, China
2 shared publications - Bingfeng Zhang
China University of Petroleum (East China), Qingdao, China
1 shared publication - Chao Yao
University of Science and Technology Beijing, Beijing, China
1 shared publication - Haocheng Li
Suzhou Dinnar Automation Co., Ltd. Suzhou, China
1 shared publication - Jiaming Liang
Institute of Information Science, Beijing Jiaotong University, Beijing, China
1 shared publication - Jian Jin
Nanyang Technological University
1 shared publication
Latest Publications
- 10 March 2026
- Published by Elsevier in Pattern Recognition
- Vol. 178
Unleashing the power of optimal head in CLIP and DINO for weakly supervised semantic segmentation
- 10 March 2026
- Published by Elsevier in Pattern Recognition
- Vol. 178
Unleashing the power of optimal head in CLIP and DINO for weakly supervised semantic segmentation
- 6 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Artificial Intelligence
- Vol. PP (8) , 1-13
FGPT: Fine-Grained Prompt Tuning for Zero-Shot Anomaly Detection
Zero-Shot Anomaly Detection (ZSAD) enables models to identify anomalies in previously unseen classes by lever-aging knowledge learned from auxiliary datasets. While current ZSAD approaches leverage large pre-trained Vision-Language Models (VLMs) for their generalization capabilities, our experimental analysis reveals inherent false activations in these methods. To address this limitation and enhance the visuallinguistic connection capabilities of VLMs in zero-shot scenarios, we introduce a novel framework called Fine-Grained Prompt Tuning (FGPT) for the ZSAD task. Our main insight is to learn patch-specific prompts, departing from the conventional approach of using a shared prompt across all patches. This strategy enables more precise measures of individual patches, significantly mitigating the false activation problem. Our framework consists of two primary components: (1) Image-level Prompt Identification, which dynamically selects appropriate prompts for each image, and (2) Patch-level Prompt Matching, which effectively aggregates prompts for individual patches. To enhance the patch-specific prompt generation, we introduce an Entropy-aware Optimal Transport (EOT) algorithm that enables more effective prompt aggregation. We validate the effectiveness through comprehensive experiments across industrial and medical datasets, demonstrating superior performance compared to existing state-of-the-art methods in ZSAD tasks. The implementation code is available at https://github.com/JasonKyng/FGPT.
- 6 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Artificial Intelligence
- Vol. PP (8) , 1-13
FGPT: Fine-Grained Prompt Tuning for Zero-Shot Anomaly Detection
Zero-Shot Anomaly Detection (ZSAD) enables models to identify anomalies in previously unseen classes by lever-aging knowledge learned from auxiliary datasets. While current ZSAD approaches leverage large pre-trained Vision-Language Models (VLMs) for their generalization capabilities, our experimental analysis reveals inherent false activations in these methods. To address this limitation and enhance the visuallinguistic connection capabilities of VLMs in zero-shot scenarios, we introduce a novel framework called Fine-Grained Prompt Tuning (FGPT) for the ZSAD task. Our main insight is to learn patch-specific prompts, departing from the conventional approach of using a shared prompt across all patches. This strategy enables more precise measures of individual patches, significantly mitigating the false activation problem. Our framework consists of two primary components: (1) Image-level Prompt Identification, which dynamically selects appropriate prompts for each image, and (2) Patch-level Prompt Matching, which effectively aggregates prompts for individual patches. To enhance the patch-specific prompt generation, we introduce an Entropy-aware Optimal Transport (EOT) algorithm that enables more effective prompt aggregation. We validate the effectiveness through comprehensive experiments across industrial and medical datasets, demonstrating superior performance compared to existing state-of-the-art methods in ZSAD tasks. The implementation code is available at https://github.com/JasonKyng/FGPT.
- 3 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Broadcasting
- Vol. 72 (1) , 247-259
SFTNet: Spatial-Frequency Transformer Network for Learned Image Compression
In learned image compression (LIC), most existing methods process spatial domain information through convolutional neural networks (CNNs) or Transformers. However, they struggle to model frequency-domain correlations, particularly high-frequency detail correlations across regions. They suffer from high-frequency detail redundancy and inefficient uniform bit allocation, neither of which adapts to regional complexity. To address these issues, we propose the Spatial-Frequency Transformer Network (SFTNet) for LIC, which models cross-domain correlations between spatial structures and frequency components to enable region-adaptive bit allocation. To model these correlations, the Frequency-aware Transformer Block (FATB) employs dual attention mechanisms to collaboratively process spatial and frequency components. Specifically, its intra-block frequency attention enhances high-frequency details within each block by using low-frequency components as local structural guidance. Meanwhile, its inter-block spatial attention captures global consistency by modeling cross-block dependencies among low-frequency components. Building on this, the Feature Re-weighting Strategy (FRS) evaluates block importance via joint analysis of spatial dependencies and high-frequency energy. This enables dynamic alignment of the spatial-frequency features modeled by FATB for adaptive bit allocation, prioritizing structurally complex regions and compressing smooth areas to reduce redundancy. Experimental results demonstrate that our SFTNet outperforms VVC by −6.55% and −6.18% in BD-Rate on Kodak and CLIC datasets while achieving better reconstruction quality of the high-frequency details compared to recent LIC methods.
- 3 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Broadcasting
- Vol. 72 (1) , 247-259
SFTNet: Spatial-Frequency Transformer Network for Learned Image Compression
In learned image compression (LIC), most existing methods process spatial domain information through convolutional neural networks (CNNs) or Transformers. However, they struggle to model frequency-domain correlations, particularly high-frequency detail correlations across regions. They suffer from high-frequency detail redundancy and inefficient uniform bit allocation, neither of which adapts to regional complexity. To address these issues, we propose the Spatial-Frequency Transformer Network (SFTNet) for LIC, which models cross-domain correlations between spatial structures and frequency components to enable region-adaptive bit allocation. To model these correlations, the Frequency-aware Transformer Block (FATB) employs dual attention mechanisms to collaboratively process spatial and frequency components. Specifically, its intra-block frequency attention enhances high-frequency details within each block by using low-frequency components as local structural guidance. Meanwhile, its inter-block spatial attention captures global consistency by modeling cross-block dependencies among low-frequency components. Building on this, the Feature Re-weighting Strategy (FRS) evaluates block importance via joint analysis of spatial dependencies and high-frequency energy. This enables dynamic alignment of the spatial-frequency features modeled by FATB for adaptive bit allocation, prioritizing structurally complex regions and compressing smooth areas to reduce redundancy. Experimental results demonstrate that our SFTNet outperforms VVC by −6.55% and −6.18% in BD-Rate on Kodak and CLIC datasets while achieving better reconstruction quality of the high-frequency details compared to recent LIC methods.