Research Areas

Machine learning and deep learning use algorithms to parse data, learn from it, and make decisions or predictions about real-world events. Unlike traditional software programs with hard-coded instructions for specific tasks, these methods are "trained" using large amounts of data, learning how to complete tasks through various algorithms.

Deep learning is a subset of machine learning, and both are methods for implementing artificial intelligence. Our team researches and develops machine learning and deep learning while solving practical problems, contributing to the advancement of artificial intelligence.

In the interdisciplinary context of computer science, materials science, and materials physics, AI-based materials research has become a strategic priority and core technology development area worldwide. In 2012, China launched the "Materials Science System Engineering Development Strategy Research—China's Materials Genome Initiative" project, which has significant strategic importance for developing materials-related theories and applications to gain advantages in future intelligent society development. However, we currently know little about the relationships between material structure, morphological characteristics, and material properties, and synthesis methods are mostly based on improvements to macroscopic materials, which has certain limitations.

The small size effect, surface effect, and quantum effect of nanomaterials lead to complex image textures, variable morphologies, and extremely difficult image processing. Currently, the main method for observing and analyzing material images is still manual, resulting in substantial waste of human resources and time. With the development of machine learning and deep learning methods, high-speed and effective automatic analysis of material images has become possible.

Based on machine learning and deep learning research, our research group has proposed various image processing methods for materials and implemented corresponding software models. These provide data support for the Materials Genome Initiative, enabling computers to automatically recognize and classify material morphological features, effectively saving time and human resources in material data analysis, accelerating research output, and promoting the development of materials science. Research and productization of automatic analysis of material image feature information will have far-reaching significance for the development of materials science and the improvement of related artificial intelligence technologies in China.

Materials text mining utilizes natural language processing and machine learning techniques to automatically extract valuable information from vast amounts of unstructured text in materials science literature, patents, and reports. By constructing materials knowledge graphs, automating literature reviews, and extracting relationships, it greatly improves the efficiency of information acquisition in materials development and accelerates the discovery of new materials and performance prediction. Our team is dedicated to developing efficient text mining algorithms and systems to enable intelligent management and in-depth utilization of materials science data.

Robots identify, select, and classify objects through observation. Traditional vision methods use the SIFT algorithm for image matching to achieve object recognition, but this approach is slow, making robot actions very sluggish. I use one-dimensional barcodes and two-dimensional QR codes to determine the positional relationship (including angle and coordinates) between objects and robots, helping robots quickly complete object matching work and implement corresponding actions.

Additionally, I also use infrared technology and infrared observation instruments, attaching special tags to specific locations so robots can determine their position. These studies are conducted within the RT-Middleware framework. RT-Middleware can control a robot by modules, allowing researchers to separate robot parts into independent modules and develop these independent modules to upgrade and develop robot functionality.

Human-computer interaction is a major topic in today's rapidly developing information industry. Gesture control's natural and convenient characteristics align with the direction and goals of human-computer interaction optimization, offering advantages that traditional interaction methods relying on mice and keyboards lack. For certain applications, while touchscreens have improved user experience, their high cost typically limits them to small mobile devices and personal computers. Gesture control simulates touchscreen effects with a larger screen control range, serving as an important means to enhance large-screen operation experience. However, the diversity of gestures themselves, coupled with interference from unstable environmental factors, determines the complexity of gesture judgment. How to track, capture, and recognize dynamic gestures in complex backgrounds in real-time, ensuring the smoothness of the interaction process, is the research question for gesture recognition-based human-computer interaction.

Image processing is a technique that uses computers to analyze images to achieve desired results. It is the foundation for computer vision and robot vision research. Object recognition is one of the most important areas of image processing research. Traditional object recognition algorithms mainly use data from image libraries to match real-world object images, determining the type of object and recognizing it. Many objects have multiple forms; for example, scissors display different shapes depending on the angle of opening. Additionally, observing from different angles results in different object shapes. We refer to the recognition of objects with different forms as RMCO (the recognition of multiple configurations of objects). The general method involves collecting multiple data points of the same object to implement RMCO, which wastes a lot of time and makes the database very large. Through shape space theory, I can predict possible shapes of observed objects with very little data, achieving recognition of multi-form objects with minimal data.

Traditional object recognition algorithms, including SIFT and SURF, can complete object recognition in a very short time, but in robot applications, they remain an important factor limiting rapid robot action. I use three invariant factors in images: pixel information around feature points, angle information of feature points, and the information that feature point groups between two images maintain relative positions unchanged to match these two images. This approach can improve speed by about 10 times compared to SIFT while maintaining almost no incorrect matching points between matching point groups.