Xufeng Lin Email & Phone Number

Sydney, New South Wales, Australia

As part of Data Science Pillar, my primary focus is on addressing the intricate challenges that arise in the realm of microscopy image analysis. By harnessing the immense power of computer vision and machine learning, we are committed to unravelling the mysteries hidden in complex biological images and ultimately paving the way for groundbreaking advancements in medical research and diagnostics.

Geelong, Victoria, Australia

1. Disinformation Detection. Researched and developed techniques for detecting facial expression manipulation and entire face synthesis by employing deep neural networks (e.g., ResNet, Xception, etc.) to learn the spatial and spectral patterns from fake facial images generated by Generative Adversarial Networks (GAN). Given the large number of GAN variants and the potential issue of "concept drift" in real-world scenarios, technologies based on Continual Learning on limited new data were also developed to pave the way towards generalizable deepfake detection.2. Smart Video Surveillance. Developed and implemented novel solutions to track multiple individuals and infer malicious behaviors and routes by analyzing the videos captured by a set of overlapping and non-overlapping FOV cameras deployed at airport terminals. It involved: 1) object detection with YOLOv4, 2) person re-identification (Re-ID) with multi-scale deep features, 3) detection to track association with hybrid track association and bipartite graph matching, 4) tracking accuracy improvement by exploiting appearance, temporal, camera network topology, and trajectory similarity information under the Bayesian information fusion framework, and 5) anomaly detection based on Decision Tree and Deep Bayesian Network for detecting suspicious routes and behaviors.3. Precision Farming. Developed and implemented computer vision algorithms to remotely monitor plant growth and gain insights from a variety of environmental data for optimizing the cultivation of cannabis. My work included: 1) developing a novel self-supervised leaf segmentation algorithm to monitor plant growth without using annotated training data, 2) applying logistic models for simulating the growth of cannabis in terms of leaf area index to identify the ideal harvest time, and 3) developing an end-to-end self-supervised color correction model based on U-Net to correct the color distortion in images captured under artificial grow lights.

Wagga Wagga, New South Wales, Australia

- Conducting research through scholarship, publication, and presentation- Collaborating on research projects with internal and external partners- Providing discipline input to computing course and subject design- Teaching and supervising undergraduate, honours, and HDR students

Reykjavík, Capital Region, Iceland

The purpose of this project is to develop innovative software services for visual imagery that helps police, security and other agencies automatically identify, analyze, and link images on websites, seized computers and storage devices. My work involved:1. Data Collection. Built an automatic data collection pipeline using Python to automatically collect more than 200,000 images from an online photo-sharing website Flicker.com based on camera models, image resolutions, and image qualities. These images were used for evaluating the performance of source-oriented image clustering and query algorithm.2. Large-Scale Image Clustering. Developed a novel divide-and-conquer based algorithm to efficiently cluster more than 1.5 million images/videos in INTERPOL’s Child Sexual Exploitation (ICSE) database. As a result, the clustering time on the ICSE database was reduced from weeks to hours while maintaining an F1 score of >75% and a precision of >95%.3. Large-Scale Image Query. Developed an efficient image query algorithm to query against the clustering results of the ICSE database for a given image. It is able to retrieve the top-K clusters that are most likely to share the same source camera with the query image and output the corresponding likelihoods. The developed algorithm allows linking the images obtained elsewhere to those in the ICSE database.4. Forensic Image Identifier. Packed the above features into C/C++ APIs and worked with colleagues to integrate them into a large-scale image/video identification system. The software was delivered to INTERPOL to identify potential child sexual abuse criminals and victims on the Internet or seized computers and storage devices.

Shenzhen, Guangdong, China

- Developing software for Oracle Enterprise Manager (OEM)- Tracking issues and version control using VCS

Doctor Of Philosophy (Ph.D.), Computer Science

Image forgery detection, source camera identification and large-scale image clustering based on device fingerprint.

Master’S Degree, Signal And Information Processing

Applications of digital forensics in image source identification and video tampering detection.

Bachelor’S Degree, Electronics And Information Engineering