Pengze Li

Text2hologram: describe your hologram here

04/2023 - 09/2023

Academic Supervisor: Dr. Kaan Aksit

The fusion of natural language processing, image processing, and holography paves the way for unprecedented innovations in the field of computer-generated holography. This dissertation presents text2hologram, a groundbreaking pipeline that leverages deep learning models to transform textual descriptions into phase-only holograms. Aimed at democratizing access to holography and stimulating further research, the pipeline features a modular architecture and offers diverse interaction modes to cater to a wide range of users. This research employs a comprehensive methodology, that includes text-to-image conversion, super-resolution, depth estimation, hologram generation, and reconstruction at specific depth levels. The pipeline's user-centered design is emphasized, highlighting its focus on ease-of-use and reproducibility. Results and sample outcomes are presented, demonstrating the pipeline's efficacy through various intermediate outputs. The dissertation also concludes with a critical discussion where the novelty of the pipeline, its limitations, and potential future directions are examined. The source code is released to both PyPI and Github platforms, and an online Colab web tutorial is presented at here

Multi-Task Learning for Segmentation of Male Pelvic Structures in MRI

03/2023 - 05/2023

Academic Supervisor: Dr Andre Altmann

This research utilizes a publicly accessible database from UCL, comprising 589 T2-weighted MR images of the male pelvic region, with eight anatomical structures segmented. Our approach incorporates multi-task learning (MTL) segmentation for precise definition of these crucial areas. This method enhances the segmentation accuracy by drawing information from adjacent structures with distinct shape characteristics. Tested on four anatomical regions with blurred boundaries and variable shapes, our MTL approach demonstrated superior accuracy, outperforming standalone tasks, the baseline U-Net, and other multitask models. The report is at here.

3-D Painting Virtual Reality Platform

02/2023 - 04/2023

Academic Supervisor: Prof. Anthony Steed

In cooperation with Lucien Li, I completed a virtual reality platform that provides the function of drawing in 3D space and supports online multiplayer collaboration. The platform is done using Unity, run in Unity or a VR headset. The platform is based on Ubiq, UCL's open-source social virtual reality platform. On the left is the perspective of the player in the virtual environment, and another player is drawing eyes on the snowman. The complete final report is at here.

Mean Curvature Skeletons

04/2023 - 05/2023

Academic Supervisor: Prof. Niloy J. Mitra

Building on the concepts in "Mean Curvature Skeletons," this study creates an open-source Python model for deriving skeletons from 3D objects. Using Voronoi poles and the Laplacian matrix, our model iteratively solves equations to approximate an object's skeleton, identifying points near this structure. The model's efficiency is enhanced through optimizations in collapse, eigenvector reconstruction, and Laplace smoothing. This method effectively determines the skeletal structure of complex objects, marking the first open-source Python implementation of this kind. The open source code is at here.

Pedestrian Classification

01/2022 - 06/2022

Academic Supervisor: Dr. Da Guo, Prof. Xile Cao

My graduation project was cooperated with the engineers from Aidong Beyond Artificial Intelligence Technology (Beijing) Co., Ltd. Taking advantage of the private dataset of pedestrians in the factory provided by Aidong, I processed real industrial datasets and completed a pedestrian classification model based on ResNet, in which data augmentation and SE Module are applied. The accuracy (TP+TN)/(TP+FN+FP+TN) is greater than 96% . Considering the needs of the industry, this project ran with TensorFlow.

Image Processing Intern

06/2021 - 06/2022

Institute of Automation,
Chinese Academy of Science

Academic Supervisor:
Prof. Jian Cheng, Dr. Chenghua Li

Job Responsibilities:
* Stay updated with the state-of-the-art computer vision and deep learning algorithms;
* Implement convolutional neural network (CNN) models on GPU, based on Python and Pytorch;
* Focus on real-time denoising algorithms, with an emphasis on U-Net-based approaches.
* Design, train, evaluate, tune, and accelerate models, one of the results is shown below.
The video shows the denoising results (25 frames) given by a convolutional neural network in 0.5 seconds.

Aerial Image Road Extraction

09/2020 - 06/2022

Academic Supervisor: Dr. Junli Yang

In the CVPR DeepGlobe Road Extraction Challenge, our team developed a novel approach called Multiscale Strip Pooling DLinkNet. We designed a fusion model incorporating various architectures, including ResNet, D-LinkNet, SENet, D-block, Transformer, and several other modules. By employing a pooling feature map to address the challenges associated with narrow and slender features in road extraction tasks, we achieved notable improvements in IoU scores. A patent for this work is currently pending (Patent Application No.: 2023106206149), and a paper detailing our findings will be submitted to the journal Remote Sensing.

National University of Singapore Online Winter Programme: Artificial Intelligence and Machine Learning

01/2021 - 02/2021

Academic Supervisor: Prof. Mehul Motani

Supervised by Assoc. Prof. Mehul Motani, I acquired foundational knowledge in supervised learning algorithms such as Linear Regression, Decision Tree, SVM, and Neural Networks. The program offered hands-on experience in utilizing Python and Jupyter Notebook, incorporating libraries such as Pandas, NumPy, and Scikit-learn. My key project during this period involved predicting the future populations of Singapore and China using the learned models, executed in a team-oriented environment.