Shoukun Sun, Ph.D.

I am a Postdoctoral Researcher in the Department of Computer Science at the University of Idaho, working with Xiaogang Ma. I build computer-vision and generative-AI methods for scientific imaging in settings where labels are scarce and images are very large: interactive segmentation that needs fewer clicks, diffusion models that generate and super-resolve images of arbitrary size, and training-free co-registration of satellite imagery.

I received my Ph.D. in Computer Science from the University of Idaho in 2025, advised by Min Xian, with a dissertation on deep learning for critical applications under sparse labeling. Much of that work was done with Idaho National Laboratory, on micrographs of irradiated nuclear fuel and on reactor event reports. The annotation platforms I built along the way, Auto-IS and SegMap, are in routine use at INL, the University of Idaho, NCEAS (UC Santa Barbara), and The Nature Conservancy.

Selected work

Auto-IS and SegMap: AI-assisted annotation platforms

Two annotation platforms built around my interactive-segmentation models, so that scientists label their own data with a few clicks instead of tracing outlines: Auto-IS for general imagery, with AI-assisted labelling, incremental training and group collaboration, and SegMap for remote sensing, as a QGIS plugin and a web app. In routine use at Idaho National Laboratory, the University of Idaho, NCEAS (UC Santa Barbara) and The Nature Conservancy, they underpin 12+ interdisciplinary projects and cut data-preparation time by up to 90%.

SegMap code·All platforms

Interactive segmentation with fewer clicks

CFR-ICL trains an interactive segmentation model with an iterative click loss, the first loss to put the number of user clicks into the training objective, and refines masks at inference by running the same network over its own output. It set the state of the art on five benchmarks and needs 33% fewer clicks than the previous best to reach 0.95 IoU on Berkeley. It has since become the annotation engine behind our TRISO nuclear-fuel datasets.

AAAI 2024·Paper·Code

Diffusion models beyond their native resolution

Averaging overlapping patches lets a small diffusion model paint a large image, but it quietly erodes the noise the sampler depends on, leaving seams and blur. Guided and variance-corrected fusion repairs this at no extra cost, taking a 512-pixel text-to-image model to seamless panoramas seven times wider and lowering FID from 11.19 to 4.02. InfScene-SR carries the same correction into super-resolution of arbitrarily large satellite scenes at constant GPU memory, within 0.003 IoU of native high-resolution imagery on a downstream mapping task.

AAAI 2025 · arXiv 2026·AAAI paper·GVCF code·InfScene-SR preprint·InfScene-SR code

Satellite image co-registration that does not fail silently

Alignment tools each assume one kind of motion and fail silently on the image pairs that break it. SCDF estimates a dense per-pixel displacement field with acceptance thresholds calibrated from the image pair itself: training-free, GPU-free, whole 8192-pixel scenes on one CPU core. On a new 584-pair benchmark over 60 sites worldwide it registers every pair without failure and lowers the best baseline's median error from 6.83 to 4.17 m.

arXiv 2026 · under review at IEEE TGRS·Preprint

Measuring fission-gas bubbles in irradiated nuclear fuel

With Idaho National Laboratory: a multitask network that separates touching fission-gas bubbles in electron micrographs of irradiated U-10Zr fuel, reaching 92% instance recall from only 827 annotated bubbles, against 54% for the thresholding it replaced. The same model then measured about 230,000 pores across an entire 6 mm fuel cross section, and the collaboration continues on TRISO particle fuel with RU-Net and the uncertainty-aware UA-Net.

Sci. Rep. 2023 · Mater. Charact. 2024, 2026 · arXiv 2026·Sci. Rep. paper·UA-Net preprint

All projects →