Classifying cats and dogs
How does a self-driving car know what a road sign looks like? Creating a successful AI like the one that operates a self-driving car is a complex process involving many different tasks, and one of the most important of those tasks is image classification.
PhD student Ying Bi is working with Professor Mengjie Zhang and Associate Professor Bing Xue in the School of Engineering and Computer Science to find new techniques for better image classification which have applications in self-driving, medical diagnosis, and more.
“Image classification involves classifying images into predefined groups according to the content in the image,” Ying says. “For example, classifying images with a dog into the ‘dog class’ and images with a cat into the ‘cat class’.”
After classifying existing images into different categories like cats or dogs, researchers can show these collections of images to AI technology. The AI then learns what a cat, for example, can look like. That way, when the AI sees a cat in real-life (or sees another image of a cat), it can identify the cat successfully. In the example of a self-driving car, the AI that runs the car could identify a cat and be programmed to avoid hitting the cat while driving the car.
Ying’s research uses a technique called genetic programming for image classification, which has benefits over other image classification techniques. Because there is so much variation in the way things can look, it’s important to keep developing new image classification techniques to make AI’s more effective at identifying different objects, like cats.
“Our project aims to develop new techniques using genetic programming to automatically learn features of images for more effective image classification,” Ying says.
Although Ying’s research won’t have any commercial applications right now, this kind of research is important to lots of commercial products—like the face recognition abilities of phones. Ying is also collaborating with a visiting researcher on using genetic programming to help with image classification of crops from remote-sensing images. Successful crop image identification is extremely useful in agriculture, Ying says.
Ying chose this research field because it is “interesting, practical, and challenging.”
“Challenges are always there—everything from stalled research to paper deadline. My supervisors often give me academic guidance and spiritual support to help me overcome these challenges, and I’ve also found a lot of support from ECRG group members and friends.”
For Ying, one of the most important parts of being a researcher is publishing papers and attending conferences.
“By connecting with other researchers, I can receive comments and feedback from researchers worldwide, which help me produce high-quality research. It also helps me keep up-to-date with state-of-the-art developments in my field, which is always fast developing.”
As an international researcher, Ying has also enjoyed living in Wellington.
“Living in Wellington is great for studying and research,” Ying says. “It’s also very peaceful, and the people are very nice—although I wish the weather was better!”
Studying a PhD has also helped Ying realise the importance of being self-motivated and goal-oriented, which helps complete the long journey that is a PhD.
Ying hopes to continue doing research in this field after graduation and keep progressing in this area.