Lohr College researcher advances AI models

Researchers in South Dakota State University's Jerome J. Lohr College of Engineering are developing algorithms that allow artificial intelligence to gather meaning from images and other data sources in support of technological advancement and scientific discovery.

As humans, when we see an object, our brains rapidly process the image and give us an immediate understanding of what it is, how far away it might be and how it relates to the world around us. Light reflects off the object and enters our eyes, where the retina converts it into electrical signals that travel to the brain. There, specialized regions of the visual cortex interpret those signals — detecting edges, motion and color; grouping shapes into recognizable forms; and drawing on memory, prior experience and context to construct meaning about the world around us.

Humans are extremely efficient at processing these electrical signals, which allows us to do highly complex tasks, like driving a vehicle on a busy highway, rather effortlessly.

Computers, on the other hand, cannot see; they can only process numbers. So how can computers complete equally complex tasks, like driving a vehicle or reconstructing a biological molecule?

A type of artificial intelligence, known as computer vision, is making this possible. In South Dakota State University's Jerome J. Lohr College of Engineering, researchers in the McComish Department of Electrical Engineering and Computer Science are playing a key role in advancing the algorithms needed for computer vision advancements.

Data from images

On the first floor of Daktronics Engineering Hall at South Dakota State University, assistant professor Ruyi Lian is developing lines of code to make computers better at seeing the world.

Her research focuses on improving the accuracy and efficiency of computer vision algorithms — some of which could one day power self-driving cars, autonomous robots and advanced virtual reality systems. While each algorithm may contain thousands of lines of code, the visible portion can appear deceptively simple.

For Lian, it might look something like this:

import cv2
from ultralytics import YOLO
model = YOLO("yolov8n.pt")
image = cv2.imread("street.jpg")
results = model(image)
results.show()

This code triggers millions — sometimes billions — of mathematical operations. Behind a single line, neural networks — a type of artificial intelligence — analyze patterns in pixel values, extract geometric features and estimate how objects exist in physical space.

Lian’s research, housed in SDSU’s 3D Perception and Scientific Imaging Laboratory, focuses on “image-based object pose estimation with deep geometric understanding.”

In simpler terms, her work develops algorithms that extract data from the pixels of an image and determine an object’s precise position and orientation in three-dimensional space. Rather than simply identifying a car on the road, her models can estimate exactly where that car is located, how far away it is, how it is rotated and how it occupies space relative to the camera. These insights are crucial for self-driving cars to safely operate within busy city streets.

A key aspect of Lian's work is geometric modeling. By embedding principles of 3D geometry into deep learning systems, her algorithms can infer depth and spatial relationships from flat, two-dimensional images — a significant advancement in AI research.

While "natural images" have been a major focus of her work, Lian is also applying her methods to scientific imaging. She is developing pose estimation tools for cryo-electron microscopy, an emerging imaging technique used to visualize proteins and other microscopic biological structures.

“This research is developing efficient pose estimation methods for cryo-electron microscopy images and facilitating the 3D reconstruction process,” Lian said. “We have designed an end-to-end learnable pipeline for orientation estimation. Our next step is to develop efficient full pose estimation and evaluate them on specific datasets.”

In cryo-electron microscopy, researchers capture thousands of 2D images of frozen molecules, each randomly oriented. Determining how each molecule was positioned when imaged is essential for reconstructing accurate 3D models.

Currently, generating high-resolution 3D molecular models can be computationally expensive and sometimes produces inconsistent or blurry results. More accurate and efficient pose estimation algorithms could streamline the reconstruction process and produce clearer structural models. These improvements could accelerate progress in drug discovery, vaccine development and understanding the molecular mechanisms underlying disease.

On March 27, the Lohr College of Engineering will host Innovate AI 2026, a one-day symposium that brings together industry leaders, researchers, educators and students to explore how artificial intelligence is revolutionizing health care and national security, sectors vital to the economic future of South Dakota and the Great Plains. The event will be held at Sanford's Event Barn. Sen. Mike Rounds, co-chair of the Senate Artificial Intelligence Caucus, is slated to deliver the opening remarks.