Experience
Current
Deep learning-based Sketch Extraction
Automatic line-art colorization is a highly-demanding research field owing to its labor-intensive workload and important role in illustration creation. Sketch extraction has been widely used to handle the lack of paired data in line art and illustration in the learning-based approaches of Automatic line-art colorization. This research will focus on how to reduce the domain gap between synthetic and real sketeches to improve the generalization performance of the current colorization approaches. Firstly, about 2500 illustration-artificial-sketeches pairs were collected, which provided the basis for supervised training and evaluation of the generated sketches. Then a sketch extraction model based on Swin-transformer and U-net like architecture is proposed, and trained on the proposed dataset. It can generate sketches that are more realistic while retaining more essential details than alternative approaches. This research will provide more accurate image pairs for future work in field of colorization, which will improve the generalization performance of colorization models.
Past Research
Deep learning-based terrorism risk prediction
Published at risk analysis
Developing effective counter-terrorism strategies necessitates accurate terrorism risk prediction, a challenge due to the spatiotemporal and interprovincial contagious characteristics of terrorism. This study introduces a novel spatiotemporal graph convolutional network (STGCN) approach, integrating long short-term memory, self-attention layers, and a one-dimensional convolutional neural network, to capture complex, multidimensional relationships among provinces and forecast daily risks. Three graph structures representing different contagious processes were constructed, enabling a comprehensive understanding of terrorism dynamics. The method's effectiveness is validated using Afghanistan terrorist attack data from 2005 to 2020, showcasing its superiority over other machine learning prediction models. Insights derived from this study highlight the importance of addressing both long-term root causes and short-term situational prevention for counter-terrorism management.
Interpretable spatial identity neural network‑based epidemic predictionPublished at Scientific Reports
This paper introduces the Interpretable Spatial IDentity (ISID) neural network for regional weekly infectious number predictions, addressing the challenges of overcomplexity and low interpretability in current epidemic prediction models. By simplifying the classical spatio-temporal identity model (STID) and retaining a spatial identity matrix, ISID learns regional contagion relationships with a lighter model structure. The SHapley Additive explanations (SHAP) method provides post-hoc interpretations of ISID's predictions with multivariate sliding-window time series data. Experimental comparisons demonstrate ISID's satisfactory performance in epidemic prediction, highlighting its focus on the most proximate and remote data in 20-step long input sequences, and less on intermediate steps. This study enhances reliable and interpretable epidemic forecasting, aiding public health experts in their management processes.
Award
JGC-S Scholarship for international student公益財団法人 日揮・実吉奨学会 2023年留学生奨学生
Intership
Algorithm intern at Sensetime ShanghaiAs a computer vision algorithm intern, I provided various technical support for smart industries. I worked on algorithm development, such as the identification of safety helmets, training models, and collaborated with the front-end. I primarily used Python and deep learning algorithms for my tasks. Through this experience, I realized the importance of industrial vision systems and how advanced AI technologies can significantly enhance on-site safety and efficiency.
Project
Website for Crop disease classificationOn this website, users can detect whether crops are diseased or not. By directly uploading an image, the probability of the disease is represented graphically using deep learning. The most unique feature of this project is that it doesn't have a backend. We adopted a model for edge devices, and all computations are performed on the user's browser. We hope this makes it easy to use even for farmers without computing resources.
