Research Projects

Current Research Projects

A Generative Model for Revitalising Characters with Decoupled Content and Style Injection

Supervisor: Dr. Yingfang Yuan, Heriot-Watt University, UK
Duration: May 2023 - Present

• Objective: To innovate a framework inspired by pictogram Chinese characters for generating artworks that seamlessly integrate customizable elements and styles into traditional characters, bridging ancient calligraphy with modern computational art.
• Methodologies:
  • Developing a novel generative adversarial network (GAN) architecture specifically designed to understand and manipulate the structural components of Chinese characters.
  • Implementing a style transfer mechanism that allows for the injection of diverse artistic styles while preserving the semantic integrity of the characters.
  • Creating a user interface that enables real-time interaction and customization of character elements and styles, enhancing the accessibility of the system for both artists and general users.
  • Utilizing advanced computational techniques, including differential rendering and neural style transfer, to achieve a harmonious synthesis of traditional calligraphy and contemporary digital art.

Global Urban Sustainable Development Strategies and Empirical Research

Program: Ural Federal University Program of Development within the Priority-2030 Program
Supervisor: Prof. Zhang Guoxing,Lanzhou University
Duration: May 2022 - June 2024

• Objective: To conduct a comprehensive analysis of the factors influencing urban green development and their impact on policy mechanisms, with the aim of informing sustainable urban planning strategies globally.
• Methodologies:
  • Employing advanced machine learning algorithms, including random forests and gradient boosting, to identify key predictors of urban sustainability across diverse global contexts.
  • Utilizing time series analysis and state-space models to discern both long-term equilibrium trends and short-term dynamics in urban green policy effectiveness.
  • Implementing geospatial analysis techniques to map and visualize the spatial distribution of sustainability indicators across urban landscapes.
  • Developing a novel index for quantifying urban green development, incorporating multidimensional factors such as environmental quality, economic efficiency, and social equity.

FPGA-Based AI Doctor: Deep Learning-Based Clinical Target Delineation for Cervical Cancer

Program: National College Student Innovation and Entrepreneurship Training Program
Supervisor: Prof. Wang XingHua,Lanzhou University
Duration: Mar 2024 - Present

• Objective: To enhance the capability of identifying subtle features in medical images for improved clinical target delineation in cervical cancer treatment, leveraging the parallel processing power of FPGA technology.
• Methodologies:
  • Refining the traditional U-Net architecture to incorporate residual connections and attention mechanisms, optimizing feature extraction in complex medical imaging scenarios.
  • Implementing a custom FPGA design to accelerate convolutional operations, achieving real-time processing of high-resolution medical images.
  • Developing a novel data augmentation pipeline specifically tailored for cervical cancer imaging, addressing the challenge of limited labeled datasets in medical AI.
  • Integrating a multi-scale feature fusion module to enhance the network's ability to detect lesions of varying sizes and shapes.

UNet-Centric MambaMorph: A Comprehensive Visual Mamba Framework Enhanced with Cross-Scan and Semi-Supervised Learning for Medical Segmentation

Program: Fundamental Research Funds for Central Universities Research Capacity Improvement Project
Supervisor: Prof. Zhang Wenting,Lanzhou University
Duration: Jan 2024 - Present

• Objective: To improve medical image segmentation accuracy by enhancing global context understanding through the integration of UNet and Mamba architectures, complemented by novel Cross-Scan modules and semi-supervised learning techniques.
• Methodologies:
  • Designing a hybrid architecture that combines the local feature extraction capabilities of UNet with the long-range dependency modeling of Mamba networks.
  • Implementing a Cross-Scan module to capture multi-directional contextual information, enhancing the network's ability to delineate complex anatomical structures.
  • Developing a semi-supervised learning framework to leverage large amounts of unlabeled medical imaging data, addressing the perennial challenge of limited annotated datasets in medical AI.
  • Utilizing curriculum learning strategies to progressively increase the complexity of training samples, optimizing the learning process for improved generalization.

Recommendation Algorithm Based on Knowledge Graph and Strong-Weak Connection Attention Mechanism

Program: Hui-Chun Chin and Tsung-Dao Lee Chinese Undergraduate Research Endowment
Supervisor: Prof. Su Wei,Lanzhou University
Duration: Mar 2023 - May 2024

• Objective: To refine existing recommendation algorithms by capturing subtle user group similarities through the integration of knowledge graphs and a novel strong-weak connection attention mechanism.
• Methodologies:
  • Constructing a comprehensive knowledge graph that encapsulates user preferences, item attributes, and contextual information, providing a rich semantic foundation for recommendations.
  • Developing a novel attention mechanism that differentiates between strong (direct) and weak (indirect) connections in the graph, allowing for more nuanced understanding of user-item relationships.
  • Implementing graph convolutional networks (GCNs) enhanced with the proposed attention mechanism to effectively propagate and aggregate information across the knowledge graph.
  • Designing a multi-task learning framework that jointly optimizes for recommendation accuracy and knowledge graph completion, enhancing the model's ability to handle sparse data scenarios.
• Key Results: The proposed algorithm achieves a 12% improvement in recommendation precision and a 18% increase in recall on benchmark datasets, with particularly significant gains observed for long-tail items.

Intelligent Cholesterol Management System

Program: IGEM Project
Supervisor: Prof. Li Xiangkai,Lanzhou University
Duration: Jan 2023 - Dec 2024

• Objective: To develop an intelligent system for oleic acid induction by engineering the FadO operator sequence, aimed at personalized cholesterol management across diverse human constitutions.
• Methodologies:
  • Employing CRISPR-Cas9 gene editing techniques to precisely modify the FadO operator sequence, fine-tuning its sensitivity to oleic acid concentrations.
  • Developing a mathematical model of the oleic acid induction pathway, incorporating stochastic differential equations to account for cellular heterogeneity.
  • Implementing a machine learning algorithm to analyze and predict the optimal induction thresholds for various human metabolic profiles, leveraging data from clinical trials and literature.
  • Designing a microfluidic device for real-time monitoring of cellular responses to oleic acid, enabling rapid iteration and optimization of the engineered system.
• Current Achievements: Successfully engineered a prototype system demonstrating a 5-fold increase in sensitivity to oleic acid compared to wild-type cells, with a response gradient calibrated to physiologically relevant concentration ranges (50-500 μM).

Tropical Linear Representation of Involute Chinese Monoids

Program: National College Student Innovation and Entrepreneurship Training Program
Supervisor: Prof. Zhang Wenting,Lanzhou University
Duration: Mar 2023 - May 2024

• Objective: To give new tropical representations of the Chinese monoid and compare the dimension sizes with the existing representations; also to give a tropical representation of the Chinese monoid with Schützenberger's involution.
• Methodologies:
  • For the Chinese monoid $Ch_n$ of finite rank $n\geq2$, by defining the mapping $\phi_{ij}$ and proving that it induces a tropical representation $\phi_n$ of $Ch_n$, further obtaining the mapping $\tilde{\phi}_{n}$ which is a faithful tropical representation of $Ch_n$ as a submonoid of $UT_{n(n - 1)}(\mathbb{T})$.
  • For the Chinese monoid $(Ch_n,^{\sharp})$ with involution, defining the mappings $\psi_{n}$ and $\xi_{n}$, and then through the injective homomorphism $\rho_{n}' = \iota' \circ \xi_{n}$ to get a faithful tropical representation of $(Ch_n,^{\sharp})$ as a submonoid of $(UT_{2n(n - 1)}(\mathbb{T}),^{D})$.
• Theoretical Contributions: Successfully presents new faithful tropical representations for Chinese monoids with and without involution, enriching understanding of their structure and properties in the context of tropical semirings and upper triangular matrices.