Computer Vision & ML Engineer (Perception Systems)

Yushuo Niu Æ 860-***-**** [ *********@*****.*** LinkedIn GitHub

Summary

• Experienced computer vision and ML engineer, specializing in integrating models into real-world use cases and automated imaging systems for perception, semantic segmentation, change detection, and data-efficient deep learning.

• Skilled in building end-to-end ML pipelines across data curation, annotation, model training, validation, error analysis, and deployment-oriented integration.

• Proficient in Python, C/C++, PyTorch, OpenCV and multi-GPU model training. Specialized in robust visual perception under viewpoint variation, illumination changes, noisy sensor inputs, and limited labeled data. Education

University of Connecticut (UConn), Ph.D. Computer Science and Engineering Storrs, CT, December 2025 University of Connecticut (UConn), M.S. Mechanical Engineering Storrs, CT, October 2016 Shandong University (SDU), B.S. Mechanical Design & Automation Shandong, China, July 2012 Technical Skills

Programming: Python, C/C++, CUDA, MATLAB, Bash

ML Frameworks: PyTorch, TensorFlow/Keras, Scikit-learn, OpenCV, NumPy, Pandas Computer Vision: Semantic/Instance Segmentation, Object Detection, Change Detection, Defect Detection, Temporal Image Analysis, Data-Efficient Learning

Architectures: U-Net variants, Siamese/Semi-Siamese Networks, Vision Transformers, ResNet, Autoencoder, Mask R-CNN, YOLO, Foundation Models

ML Development: Dataset Curation, Label Generation, Data Augmentation, Model Training and Validation, Error Analysis

Systems & Deployment: Linux, Git, Docker, AWS, TensorFlow Lite, ONNX Experience

Semi-Siamese Neural Networks for Real-Time Defect Detection in 3D Printing Storrs, CT Research Assistant Jun 2019 – Sep 2023

• Developed a Semi-Siamese visual perception framework for binder-jet 3D printing by comparing reference print schematics with in-situ camera images, enabling pixel-level semantic localization of no-defect, over-extrusion, and under-extrusion regions.

• Designed a cross-domain change detection model with asymmetric encoders and a shared decoder to handle large appearance gaps between schematic inputs and real camera observations.

• Built a data-efficient training and validation pipeline using only 57 experimental image pairs, combining transfer learning and camera-setup augmentation to improve robustness under viewpoint and illumination variations.

• Integrated the model into a layer-wise real-time inspection workflow, achieving 0.419 s/layer inference for real-time monitoring and corrective decision-making under practical system constraints.

• Evaluated cross-domain robustness against transformer and GAN baselines (BIT, DTCDN) on heterogeneous Optical/SAR data, reducing training time by 54% compared with BIT. MultiTaskDeltaNet: Temporal Visual Segmentation from Sequential Imaging Storrs, CT Research Assistant Apr 2023 – Jul 2025

• Designed MultiTaskDeltaNet (MTDN), a Siamese U-Net-based perception model that reformulates segmentation across sequential image frames as a temporal change detection problem, enabling robust detection of evolving semantic structures over time.

• Built a data-efficient temporal training pipeline that expanded 126 labeled frames into 2,968 paired samples, improving robustness across structural variation, imaging direction changes, and sample drift.

• Developed model validation and ablation pipelines to evaluate single-task vs. multi-task learning, focal loss, backbone initialization, and prediction fusion strategies.

• Improved hard-to-detect hollow-core regions over a U-Net baseline by +10.22% macro-F1 and +12.34% IoU, and implemented sequential inference for real-time analysis of dynamic visual changes. Automated Visual Monitoring for Self-Driving Labs Storrs, CT / Toronto, ON Research Assistant Apr 2024 – Present

• Developed a Semi-Siamese visual segmentation framework for fluorescence-free live-cell monitoring using paired brightfield and phase-contrast microscopy images, improving robustness under real-world imaging artifacts such as halos, blur, and uneven illumination.

• Built a data curation and pseudo-labeling pipeline using Cellpose 3.0 labels generated from FITC fluorescence images, enabling model training on fluorescence-free inputs while reducing manual annotation effort.

• Created preprocessing and augmentation workflows including patch extraction, flipping, rotation, cropping, blurring, and color jitter, improving generalization from low-confluency training samples (5–20%) to higher-confluency test samples

(50–70%).

• Improved over a U-Net baseline by +3.6% mF1 and +5.8% mIoU, with reduced false positives and improved detection of fine cellular structures under limited-data conditions.

• Integrated the pipeline into an automated microscopy platform (Cytation C10) within a self-driving lab workflow, supporting continuous image-based monitoring and system-level collaboration with interdisciplinary research teams. DeltaSegment: Data-Efficient Semantic Segmentation via Change Detection Storrs, CT Research Assistant May 2025 – Apr 2026

• Proposed DeltaSegment, a lightweight Semi-Siamese segmentation framework that reformulates semantic segmentation as a multifidelity change detection problem under limited annotation.

• Designed a multifidelity learning strategy by pairing each high-fidelity image with a low-fidelity, color-quantized counterpart generated from the same image, improving segmentation consistency and boundary delineation without additional data collection.

• Evaluated the framework on public segmentation benchmarks including GlaS, MoNuSeg, and TNBC, outperforming transformer-based baselines such as TransUNet and Swin-UNet with gains up to +1.5% Dice and +2.4% IoU. Blood Cell Instance Segmentation with a Pathology Foundation Encoder Storrs, CT Independent Research Project Jan 2026 – Apr 2026

• Developed an instance segmentation workflow for blood cell images by converting semantic masks into instance-level supervision using watershed and connected-components based separation for touching cells.

• Fine-tuned UNI2h, a ViT-based pathology foundation model, with a U-Net-style decoder on the BCCD dataset for cell-level segmentation.

• Compared foundation-model segmentation outputs with Cellpose-SAM-style results, supporting transferability of pathology foundation representations to biomedical cell segmentation workflows. Object Detection on Chest X-Rays with Mask R-CNN and YOLOv3 Storrs, CT Research Assistant Aug 2019 – Dec 2019

• Curated and preprocessed chest X-ray datasets through normalization, resizing, and artifact handling, then trained customized Mask R-CNN and YOLOv3 models for pneumonia detection.

• Converted model detections into visual region highlights for review, improving interpretability of detection outputs and achieving 0.73 F1 on held-out evaluation data. Awards

Predoctoral Honorable Mention, School of Computing, UConn. (May 2025) NSF I-Corps Northeast Certificate, UConn Propelus Program. (July 2025) CSE First Place Award, UConn Graduate Poster Competition. (Feb 2024) Selected Presentations

Niu, Y., et al., August 2025. Towards Autonomous Image-Based Cell Culture Analysis with a Human-Informed Semi-Siamese Neural Network Workflow. Poster presentation, Acceleration Conference 2025, Toronto, ON, Canada.

Niu, Y., et al., June 2024. Robust Defect Detection for Binder Jet 3D Printing with Semi-Siamese Neural Networks. Contributed talk, AI Innovations, TechConnect World Innovation 2024, Washington, DC. Niu, Y., et al., February 2024. Robust Defect Detection for Binder Jet 3D Printing with Semi-Siamese Neural Networks. Invited Talk, Nondestructive Inspection for AM parts (NDIxAM) 2024, New Jersey, NJ. Publications

Niu, Y., Chadwick, E., Ma, A.W. and Yang, Q., 2023, September. Semi-Siamese Network for Robust Change Detection Across Different Domains with Applications to 3D Printing. In International Conference on Computer Vision Systems (pp. 183–196). Cham: Springer Nature Switzerland.

Niu, Y., Li, T., Zhu, Y. and Yang, Q., 2025. MultiTaskDeltaNet: Change Detection-based Image Segmentation for Operando ETEM with Application to Carbon Gasification Kinetics. Digital Discovery, Royal Society of Chemistry. https://doi.org/10.1039/D5DD00333D

Niu, Y., Reed, H. and Yang, Q., 2026, April. DeltaSegment: Multifidelity Medical Image Segmentation via Change Detection. In 2026 IEEE 23rd International Symposium on Biomedical Imaging (ISBI) (pp. 1–5). IEEE. Sahoo, S., Shende, C., Hossain, M. Z., Patel, P., Niu, Y., Wang, X., Ware, S., Bi, J., Kamath, J., Russell, A., and Song, D. (2025). Cross-platform Prediction of Depression Treatment Outcome Using Location Sensory Data on Smartphones. ACM Transactions on Interactive Intelligent Systems (TiiS), 15(2), Article 27. Moaddel, T., Bertola, G., Quan, C., Dasgupta, B., Ghatlia, N., Shiloach, A., Ma, A., Niu, Y., Vinski, P. and Yang, Q., 2024. 51443 Sub-micron emulsion based mild bodywash formulations for superior active deposition. Journal of the American Academy of Dermatology, 91(3), p.AB318.

Contact this candidate