Boyuan Peng | Academic Homepage

Boyuan Peng

I received my M.S. in Electronic and Information Engineering from Tsinghua University in 2025, advised by Prof. Peiwu Qin. Previously, I earned my B.E. in Electronic and Information Engineering at China University of Mining and Technology.

My research connects multimodal AI, generative modeling and computational imaging, including audio-visual perception, physics-guided image restoration and reliable biomedical AI. My latest project, MundiSense v1, brings voice and facial expression together as Mundiscope’s first perception layer for human-centered interaction, with explicit uncertainty.

Email: pengby99@outlook.com / CV / LinkedIn Google Scholar / Github

News

Oct. 2026 MundiSense v1 introduces Mundiscope’s audio-visual perception layer. Explore the 1 min 44 s recorded demo, technical report and research code.

Jun. 2025 I completed my M.S. in Electronic and Information Engineering at Tsinghua University.

Sep. 2024 Our paper on cell segmentation under optical aberrations was published in CSBJ.

Education

	Tsinghua University 2022.09 - 2025.06 M.S. in Electronic and Information Engineering Research Advisor: Prof. Peiwu Qin
	China University of Mining and Technology 2017.09 - 2021.06 B.E. in Electronic and Information Engineering

Research & Projects

† Equal contribution. The latest project is highlighted.

My work spans multimodal LLMs, generative modeling and reliable AI for scientific and medical applications. See also Google Scholar.

Demo excerpt · recorded replay	MundiSense v1: Audio-visual emotion perception Boyuan Peng · Mundiscope · 洞悉尘世 Project / Demo / Report / Code 2026 · Research release A first sense for human interaction. Bringing voice and facial expression together, with room for uncertainty. MundiSense v1 is Mundiscope’s Sense layer for Chinese-language, single-person interaction: a LoRA-adapted Qwen2.5-Omni-7B combines with HuBERT and CLIP specialists to estimate observable expression, route missing inputs and abstain when evidence is uncertain. Sense → Remember → Express is the broader direction; memory and expressive response remain future work. Multimodal LLMs Affective Computing Audio-Visual Fusion Uncertainty-Aware AI EmotionTalk R3: macro-F1 0.645 on 1,929 utterances, without threshold rejection. CH-SIMS v2: sentiment Pearson r = 0.792 on 976 media-available clips of 1,034. These measure different tasks; expression estimates do not establish a person’s inner state.
	Practical guidelines for cell segmentation models under optical aberrations in microscopy Boyuan Peng†, Jiaju Chen†, P. Bilha Githinji†, Ijaz Gul, et al. Paper / PDF / Code 2024 · Computational and Structural Biotechnology Journal A robustness benchmark for cell instance segmentation under simulated optical aberrations in fluorescence and bright-field microscopy. The study compares segmentation architectures and Cellpose 2.0, and introduces PLCM to identify aberration properties from point spread functions. Robustness Benchmarking Distribution Shift Cell Instance Segmentation Optical Priors DynamicNuclearNet + LIVECell · Optical priors · Model-selection guidelines
	PSF-guided conditional diffusion for microscopy aberration correction Boyuan Peng, Jiaju Chen, et al. Research project Exploring point spread functions as physical conditioning for diffusion-based restoration of aberrated microscopy images. The goal is to recover cellular structure and boundaries while incorporating knowledge of the optical imaging process. Diffusion Models Physics-Guided Generative AI Conditional Image Restoration Physical embedding · Conditional diffusion · Microscopy restoration
$Illustration of coaxial ophthalmic optics, fundus and refractive imaging, alignment and image fusion.$	Dual-Modality Computational Ophthalmic Imaging with Deep Learning and Coaxial Optical Design Boyuan Peng†, Jiaju Chen†, Yiwei Zhang†, Cuiyi Peng, et al. Preprint / PDF 2025 · arXiv preprint A coaxial optical platform combines fundus imaging and refractive sensing. Learning-based pupil localization supports alignment, while retinal vessel segmentation and multimodal learning connect optical measurements with ophthalmic screening research. Multimodal Learning Medical Image Segmentation Learning-Assisted Optics Coaxial optics · Automated alignment · Retinal vessel analysis
	Convolutional neural networks for rapid diagnosis and lesion detection of pediatric mycoplasma pneumoniae pneumonia using chest X-rays Jiaming Deng†, Danqing Yu†, Boyuan Peng†, Dongmei Yu†, et al. PDF Research manuscript ConvNeXt-Tiny distinguishes mycoplasma, bacterial and viral pneumonia from normal pediatric chest X-rays. Score-CAM highlights regions associated with model predictions, with an Android application prototype for exploring results. Explainable AI ConvNeXt Medical Image Classification Model Attribution Reported test accuracy: 88.20% · AUROC: 0.9218 · Four-class research evaluation
	Intelligent proximate analysis of coal based on near-infrared spectroscopy Weinan Liu, Boyuan Peng, Xiaoyu Liu, et al. Paper / PDF 2021 · Journal of Applied Spectroscopy Near-infrared spectra provide a rapid, nondestructive route to estimating moisture, volatile matter and calorific value. The study compares random forests, extreme learning machines and particle-swarm-optimized ELM regression. Scientific Machine Learning Spectral Regression Swarm Optimization Near-infrared spectroscopy · PSO–ELM · Material characterization