Songtao Wu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Songtao Wu has a total of 35 co-authors. Here are the top co-authors by number of shared publications:
- Kuanhong Xu
Research and Development Center China, SONY Group Corporation, Beijing, China
4 shared publications - Qiang Tong
Research and Development Center China, SONY Group Corporation, Beijing, China
4 shared publications - Sheng-Hua Zhong
College of Computer Science and Software Engineering, Shenzhen University, Shenzhen, Guangdong, China
4 shared publications - Yan Liu
Department of Computer Science, Hong Kong Baptist University, Kowloon, People’s Republic of China
4 shared publications - Chen Sun
Research and Development Center Sony (China) Ltd., Beijing, China
3 shared publications - Hui Fang
Loughborough University, Loughborough, United Kingdom
2 shared publications
Latest Publications
- 13 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on AgriFood Electronics
- Vol. 4 (1) , 383-394
Tiny-PMDet: A Knowledge-Distilled Lightweight Edge-AI Detector for Grapevine Powdery Mildew Detection on CMOS Sensor Based AIoT Systems
Edge-AI-powered Artificial Intelligence of Things (AIoT) technologies enable low-power and efficient solutions for intelligent crop monitoring. This article presents Tiny-PMDet, a lightweight yet high-accuracy detector designed for on-sensor grapevine powdery mildew detection using an edge-AI-enabled CMOS (Sony IMX500) based camera. The proposed model integrates an EfficientNet-Lite0 backbone, a sBiFPN neck, and a lightweight decoupled detection head, achieving an effective balance between detection accuracy and computational efficiency. To further enhance detection precision without increasing inference cost, a loss optimization strategy combining Weighted Cross-Entropy, Efficient IoU, and Quality Cross-Entropy is adopted, together with a dual knowledge distillation scheme based on Pearson correlation and cross-head guidance. Experimental results demonstrate that Tiny-PMDet achieves 95.1% AP@0.5 and 94.4% F1-score, outperforming state-of-the-art lightweight detectors while maintaining a lightweight model size of 2.9 MB and a low computational cost of 0.9 GFLOPs. After posttraining quantization and deployment on the IMX500 sensor, Tiny-PMDet achieves real-time inference at 36 FPS with only an approximately 1% accuracy degradation. These results highlight the advantages of edge-AI CMOS sensor-based solutions for enabling low-power, on-site, and network-independent agricultural monitoring, marking a significant step toward scalable AIoT-driven smart agriculture.
- 13 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on AgriFood Electronics
- Vol. 4 (1) , 383-394
Tiny-PMDet: A Knowledge-Distilled Lightweight Edge-AI Detector for Grapevine Powdery Mildew Detection on CMOS Sensor Based AIoT Systems
Edge-AI-powered Artificial Intelligence of Things (AIoT) technologies enable low-power and efficient solutions for intelligent crop monitoring. This article presents Tiny-PMDet, a lightweight yet high-accuracy detector designed for on-sensor grapevine powdery mildew detection using an edge-AI-enabled CMOS (Sony IMX500) based camera. The proposed model integrates an EfficientNet-Lite0 backbone, a sBiFPN neck, and a lightweight decoupled detection head, achieving an effective balance between detection accuracy and computational efficiency. To further enhance detection precision without increasing inference cost, a loss optimization strategy combining Weighted Cross-Entropy, Efficient IoU, and Quality Cross-Entropy is adopted, together with a dual knowledge distillation scheme based on Pearson correlation and cross-head guidance. Experimental results demonstrate that Tiny-PMDet achieves 95.1% AP@0.5 and 94.4% F1-score, outperforming state-of-the-art lightweight detectors while maintaining a lightweight model size of 2.9 MB and a low computational cost of 0.9 GFLOPs. After posttraining quantization and deployment on the IMX500 sensor, Tiny-PMDet achieves real-time inference at 36 FPS with only an approximately 1% accuracy degradation. These results highlight the advantages of edge-AI CMOS sensor-based solutions for enabling low-power, on-site, and network-independent agricultural monitoring, marking a significant step toward scalable AIoT-driven smart agriculture.
- 11 December 2025
- Published by Elsevier in Expert Systems with Applications
- Vol. 305
UniStyleDiff: A unified diffusion-driven framework for image and video style transfer
- 11 December 2025
- Published by Elsevier in Expert Systems with Applications
- Vol. 305
UniStyleDiff: A unified diffusion-driven framework for image and video style transfer
- Open Access
ProDehaze: Prompting Diffusion Models Toward Faithful Image Dehazing
Recent approaches using large-scale pretrained diffusion models for image dehazing improve perceptual quality but often suffer from hallucination issues, producing unfaithful dehazed image to the original one. To mitigate this, we propose ProDehaze, a framework that employs internal image priors to direct external priors encoded in pretrained models. We introduce two types of \textit{selective} internal priors that prompt the model to concentrate on critical image areas: a Structure-Prompted Restorer in the latent space that emphasizes structure-rich regions, and a Haze-Aware Self-Correcting Refiner in the decoding process to align distributions between clearer input regions and the output. Extensive experiments on real-world datasets demonstrate that ProDehaze achieves high-fidelity results in image dehazing, particularly in reducing color shifts. Our code is at https://github.com/TianwenZhou/ProDehaze.
Open AccessProDehaze: Prompting Diffusion Models Toward Faithful Image Dehazing
Recent approaches using large-scale pretrained diffusion models for image dehazing improve perceptual quality but often suffer from hallucination issues, producing unfaithful dehazed image to the original one. To mitigate this, we propose ProDehaze, a framework that employs internal image priors to direct external priors encoded in pretrained models. We introduce two types of \textit{selective} internal priors that prompt the model to concentrate on critical image areas: a Structure-Prompted Restorer in the latent space that emphasizes structure-rich regions, and a Haze-Aware Self-Correcting Refiner in the decoding process to align distributions between clearer input regions and the output. Extensive experiments on real-world datasets demonstrate that ProDehaze achieves high-fidelity results in image dehazing, particularly in reducing color shifts. Our code is at https://github.com/TianwenZhou/ProDehaze.
- 29 November 2024
- Published by Springer Nature
- p. 252-268
Towards Compact Reversible Image Representations for Neural Style Transfer
- 29 November 2024
- Published by Springer Nature
- p. 252-268
Towards Compact Reversible Image Representations for Neural Style Transfer
- 28 March 2024
- Published by Elsevier in Neural Networks
- Vol. 175, 106281
Frequency compensated diffusion model for real-scene dehazing
- 28 March 2024
- Published by Elsevier in Neural Networks
- Vol. 175, 106281
Frequency compensated diffusion model for real-scene dehazing
- Open Access
- 30 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Open Journal of the Communications Society
- Vol. 5, 38-50
AI Model Selection and Monitoring for Beam Management in 5G-Advanced
This paper investigates the application of artificial intelligence (AI) to wireless technology, specifically in the context of beam management (BM) in the advanced 5th-generation (5G) communication system. Our focus lies in aligning our study with the ongoing discussions within the Third Generation Partnership Project (3GPP) as of December 2022. Instead of evaluating the performance of specific AI models, we take user equipment (UE) receiver (Rx) beam prediction as an illustrative example of AI-based BM. We explore various aspects of AI model management, including model selection, monitoring, and activation/deactivation operations, from a 3GPP perspective. For model selection, we propose deploying distinct AI models for different propagation environments, categorized based on base station (BS) transmitter (Tx) beam measurement results. Reference Signal Received Power (RSRP) serves as a pivotal key performance index (KPI) for model performance monitoring. Our simulation results indicate that, instead of training one all-encompassing AI model with numerous layers for universal application, transitioning between domain-specific AI models with fewer layers yields superior performance. Model activation/deactivation procedures determine whether AI-based BM or traditional BM should be employed in a given scenario. We also introduce the use of AI for predicting the performance of both AI-based BM and traditional BM. By comparing the performance of these strategies, we can ascertain whether link performance degradation results from AI output errors or UE movement into challenging propagation environments. This approach enables the effective management of model switching between AI-based BM and traditional BM. The simulation shows that we can reduce the number of unnecessary switches by 10%.
- 30 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Open Journal of the Communications Society
- Vol. 5, 38-50
Open AccessAI Model Selection and Monitoring for Beam Management in 5G-Advanced
This paper investigates the application of artificial intelligence (AI) to wireless technology, specifically in the context of beam management (BM) in the advanced 5th-generation (5G) communication system. Our focus lies in aligning our study with the ongoing discussions within the Third Generation Partnership Project (3GPP) as of December 2022. Instead of evaluating the performance of specific AI models, we take user equipment (UE) receiver (Rx) beam prediction as an illustrative example of AI-based BM. We explore various aspects of AI model management, including model selection, monitoring, and activation/deactivation operations, from a 3GPP perspective. For model selection, we propose deploying distinct AI models for different propagation environments, categorized based on base station (BS) transmitter (Tx) beam measurement results. Reference Signal Received Power (RSRP) serves as a pivotal key performance index (KPI) for model performance monitoring. Our simulation results indicate that, instead of training one all-encompassing AI model with numerous layers for universal application, transitioning between domain-specific AI models with fewer layers yields superior performance. Model activation/deactivation procedures determine whether AI-based BM or traditional BM should be employed in a given scenario. We also introduce the use of AI for predicting the performance of both AI-based BM and traditional BM. By comparing the performance of these strategies, we can ascertain whether link performance degradation results from AI output errors or UE movement into challenging propagation environments. This approach enables the effective management of model switching between AI-based BM and traditional BM. The simulation shows that we can reduce the number of unnecessary switches by 10%.
- Open Access
- 26 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 2856-2864
Dynamic Compositional Graph Convolutional Network for Efficient Composite Human Motion Prediction
With potential applications in fields including intelligent surveillance and human-robot interaction, the human motion prediction task has become a hot research topic and also has achieved high success, especially using the recent Graph Convolutional Network (GCN). Current human motion prediction task usually focuses on predicting human motions for atomic actions. Observing that atomic actions can happen at the same time and thus formulating the composite actions, we propose the composite human motion prediction task. To handle this task, we first present a Composite Action Generation (CAG) module to generate synthetic composite actions for training, thus avoiding the laborious work of collecting composite action samples. Moreover, we alleviate the effect of composite actions on demand for a more complicated model by presenting a Dynamic Compositional Graph Convolutional Network (DC-GCN). Extensive experiments on the Human3.6M dataset and our newly collected CHAMP dataset consistently verify the efficiency of our DC-GCN method, which achieves state-of-the-art motion prediction accuracies and meanwhile needs few extra computational costs than traditional GCN-based human motion methods.
- 26 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 2856-2864
Open AccessDynamic Compositional Graph Convolutional Network for Efficient Composite Human Motion Prediction
With potential applications in fields including intelligent surveillance and human-robot interaction, the human motion prediction task has become a hot research topic and also has achieved high success, especially using the recent Graph Convolutional Network (GCN). Current human motion prediction task usually focuses on predicting human motions for atomic actions. Observing that atomic actions can happen at the same time and thus formulating the composite actions, we propose the composite human motion prediction task. To handle this task, we first present a Composite Action Generation (CAG) module to generate synthetic composite actions for training, thus avoiding the laborious work of collecting composite action samples. Moreover, we alleviate the effect of composite actions on demand for a more complicated model by presenting a Dynamic Compositional Graph Convolutional Network (DC-GCN). Extensive experiments on the Human3.6M dataset and our newly collected CHAMP dataset consistently verify the efficiency of our DC-GCN method, which achieves state-of-the-art motion prediction accuracies and meanwhile needs few extra computational costs than traditional GCN-based human motion methods.
- 20 July 2023
- Published by Springer Nature in Signal, Image and Video Processing
- Vol. 17 (8) , 4199-4207
A real-time detector of chicken healthy status based on modified YOLO
- 20 July 2023
- Published by Springer Nature in Signal, Image and Video Processing
- Vol. 17 (8) , 4199-4207
A real-time detector of chicken healthy status based on modified YOLO
- 1 September 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Vehicular Technology Magazine
- Vol. 17 (3) , 26-33
User Selection for Federated Learning in a Wireless Environment A Process to Minimize the Negative Effect of Training Data Correlation and Improve Performance
- 1 September 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Vehicular Technology Magazine
- Vol. 17 (3) , 26-33
User Selection for Federated Learning in a Wireless Environment A Process to Minimize the Negative Effect of Training Data Correlation and Improve Performance
- 3 June 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 22 (1) , 256-270
A Novel Convolutional Neural Network for Image Steganalysis With Shared Normalization
Image steganalysis is to discriminate innocent images (cover images) and those suspected images (stego images) with hidden messages. The task is challenging since modifications to cover images due to message hiding are extremely small. To handle this difficulty, modern approaches proposed to use Convolutional Neural Network (CNN) models to detect steganography with paired learning, i.e. cover images and their stegos are both in training set. In this paper, we explore an important technique in CNN models, the Batch Normalization (BN), for the task of image steganalysis in the paired learning framework. Our theoretical analysis shows that a CNN model with multiple batch normalization layers is difficult to be generalized to new data in the test set when it is well trained with paired learning. To address this problem, we propose a novel normalization technique called Shared Normalization (SN) in this paper. Unlike the BN layer utilizing the mini-batch mean and standard deviation to normalize each input batch, SN shares consistent statistics for training samples. Based on the proposed SN layer, we further propose a novel neural network model for image steganalysis. Extensive experiments demonstrate that the proposed network with SN layers is stable and can detect the state-of-the-art steganography with better performances than previous methods.
- 3 June 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 22 (1) , 256-270
A Novel Convolutional Neural Network for Image Steganalysis With Shared Normalization
Image steganalysis is to discriminate innocent images (cover images) and those suspected images (stego images) with hidden messages. The task is challenging since modifications to cover images due to message hiding are extremely small. To handle this difficulty, modern approaches proposed to use Convolutional Neural Network (CNN) models to detect steganography with paired learning, i.e. cover images and their stegos are both in training set. In this paper, we explore an important technique in CNN models, the Batch Normalization (BN), for the task of image steganalysis in the paired learning framework. Our theoretical analysis shows that a CNN model with multiple batch normalization layers is difficult to be generalized to new data in the test set when it is well trained with paired learning. To address this problem, we propose a novel normalization technique called Shared Normalization (SN) in this paper. Unlike the BN layer utilizing the mini-batch mean and standard deviation to normalize each input batch, SN shares consistent statistics for training samples. Based on the proposed SN layer, we further propose a novel neural network model for image steganalysis. Extensive experiments demonstrate that the proposed network with SN layers is stable and can detect the state-of-the-art steganography with better performances than previous methods.