Proposed a text-to-video generation method based on diffusion model. This method can achieve zero-shot video generation without any training by just adjusting the existing diffusion model.

• Used cross-frame attention mechanism to maintain the consistency of scene and background in the generated video, improving the video quality and coherence
• Used ControlNet to control the generation process. It can control the shape, position, size and other attributes of objects in the video by using edge, pose, semantic mask, image depth and other ways
• Highlight Contribution - Implemented edge deployment on NVIDIA Jetson Orin NV development board and built a front-end interactive page using Gradio, demonstrating the feasibility and effect of the model on low-power devices

As the main person in charge of the alumni data platform project in our department, I was responsible for requirement analysis, solution design, data processing, platform development and testing. Using Feishu system as a data storage and visualization tool, I successfully cleaned, integrated and standardized the original data, eliminating errors, omissions and inconsistencies. I created a user-friendly website using Vue framework. The website provides some special features, such as batch import/export, data analysis and statistics. The final product received high praise from our teachers and alumni, achieving data compatibility, relevance, sharing, security and low development maintenance cost. It now has nearly 100,000 alumni data collected from the founding of the school to the present.

• Highlight Contribution - Developed front-end interactive interface for alumni data platform project, using Vue + Flask to complete image upload, image annotation, model prediction, feedback and other functions.

Built a highly generalizable model and performed transfer learning to cope with the challenges of different datasets from different institutions, different annotation methods, and different datasets in the competition.

• Used Resnext-50, SegFormer, mit-b2, coat and other models for multi-level feature extraction and semantic segmentation
• Used various data augmentation methods such as Cutmix, staining and normalization tools to expand the dataset, as well as manual annotation.
• Used various techniques and innovations such as manual annotation of lung dataset, staining and normalization tools, multiple initialization, pseudo-labeling etc. to improve the accuracy and efficiency of the model
• Used pseudo-labeling. TTA, threshold adjustment, ensemble learning and other methods to optimize the model parameters and results
• Improved from 0.68 to 0.82 on the private leaderboard similarity score, ranking 20/1175 (as of September 12th, 2022)

In the summer research project at NTU, for the pneumonia CT medical image detection problem, used ResNet model to identify and detect pneumonia symptoms, and used Grad-CAM and Guided Grad-CAM methods to generate attention heat maps and explain the model’s decision process, achieving 90% accuracy

• Highlight Contribution - Responsible for the development of the front-end interactive interface, using Vue + Flask to complete the functions of image upload, image annotation, model prediction, result feedback and so on