 |
Zhen HeAssociate ProfessorDepartment of Computer Science and Computer Engineering La Trobe University Bundoora, Victoria 3086 Australia Tel : + 61 3 9479 3036 Email: z.he@latrobe.edu.au Building: Beth Gleeson, Room: 235 |
Research
My research interests are in deep learning and AI. I lead a deep learning research group at La Trobe university. There are currently 5 students in the group working on several different projects. Since 2014 my research group has been primarily focused on the application of deep learning and AI algorithms to solve a range of computer vision problems. Our main interest is in building deep learning based software to make real world impact. So we like building real systems that are deployed into the real world to automatically annotate data in a scalable way.Our most significantly contribution is the development of the SPARTA 2 system for swimming competition analysis with industry partners the Australian Institute of Sports and Swimming Australia. The system used deep learning algorithms to track swimmers and detects the start of swim strokes. This information is of vital importance to swim coaches, since it is one of the main basis for decision making on swimming tactics during swimming competitions. Traditionally this information was collected manually by many data analysts who work deep into the night during swimming competitions. Now using our deep learning software all the competition analysis information is automatically extracted by two high powered laptop computers. SPARTA is likely the most advanced competition analysis software in the world built using AI. SPARTA 2 was used to perform all competition analysis during the Tokyo Olympics in 2021. It ran flawlessly during the Olympics and producing analysis in near real time.
Our research group also built the prototype for an automated diving classification system which has been made into a production system used for tracking the load of Australian Olympic divers. The system used a deep learning algorithm to first detect the start and end of a dive and then detect the diver in the shot and finally classified the type of dive.
We have also worked with the largest telcommunication company in Australia, Telstra on a system that assists customer service agents to more rapidly respond to customer requests during online chat sessions. The system uses deep learning algorithms to make recommendations to the customer service agents and has the system was put into production with more than 1500 customer service agents using it.
Our research group is currently working in three major research areas. 1) The application of deep learning to digital pathology to help pathologists detect cancer faster and more accurately from tissue samples. 2) The detection of alcohol content from images, text and videos. 3) The use of deep learning for precision agriculture.
For the first major research area we are working with all the best cancer research institutions in Melbourne including Olivia Newton John Cancer Wellness & Research Centre, Walter Eliza Hall Institute of Medical Research, Peter MacCallum Cancer Centre and Austin Hospital. In total we are working with 5 oncologists and 1 pathologist on a large collection of scanned tissue slides consisting of more than 5000 whole slide images (WSIs). Using deep learning for digital pathology is challenging since the input data often consists of WSIs which can often be 40, 000 x 40, 000 pixels in size. The algorithm needs to find signs of cancer among the huge number of pixels. More specifically we have projects in the following areas: detecting various biomarkers such as scoring Tumour-Infiltrating Lymphocytes (TIL) and PD-L1, classifying various cancer types on WSI, predicting the survival of cancer patients from WSI and predicting various kinds of genetic mutations from WSI. In order to perform these tasks we need to perform the following computer vision tasks: segmentation, object detection and image classification.
For the 2nd major research area we are working with the Center for Alcohol Policy Research at La Trobe University. We have won several major grants focused on measuring the prevelence of alcohol on various kinds of digital media including social media, movies, song lyrics, etc. Many existing research have established a “clear link” between alcohol exposure and alcohol use, hence it is really important to know how much alcohol images, text and video people are exposed to. This could help policy makers decide how to moderate the alcohol content that people are exposed to.
For the 3rd major research area we are working with the several collegues from the Australian National University, CSIRO and Sydney University on various uses of deep learning for precision agriculture. The majority of the projects are focused on using hyperspectrum images to predict various physiological traits for wheat. Photosynthesis and related traits such as leaf area, leaf dry mass per area, are very useful for selection of superior germplasm in breeding programs. However, these traits are very expensive and time intensive to measure. In contrast, taking hyperspectrum measurements is much cheaper and faster. Hence using drones powered with hyperspectrum cameras to capture hyperspecturm images and then feeding it as input to deep learning models to predict physiological traits is a much more scalable way to select superior germplasm over large field trials. We are also working on predicting various useful plants traits from satelliate images.
Due to the very practical real world problems that we work on students have acquired very industry desirable skills. Hence former students have found jobs in places like, Nvidia, Telstra, Zendesk, CashApp, National Australia Bank, AGL, etc.
We are very actively recruiting Masters/Honours and PhD students to work on the above areas. If you are interested in enrolling into a PhD program or Masters/Honours thesis project with me, please contact me via email.
List of Research Interests
- Deep Learning
- Computer Vision
- Natural Language Processing