Yufei Hu

PhD Candidate Applicant in Computer Science & Multimodal AI
Institut Polytechnique de Paris | ENSTA Paris, France

I have achieved my Master’s degree in Engineering from École Nationale Supérieure de Techniques Avancées (ENSTA Paris), and my Bachelor’s degree in Automation from Chang'an University, Xi'an, China.

Currently, I am working at CVTE, a publicly listed company, where I have been engaged in algorithm design including machine learning, deep learning, and image processing for approximately two years.

My research focuses on advancing Multimodal AI systems, with an emphasis on integrating vision-language architectures, optimizing edge computing solutions, and leveraging self-supervised learning techniques. I am particularly passionate about bridging theoretical advancements in AI with practical, real-world applications—especially in industrial automation and resource-constrained edge environments.

I am always open to inquiries and collaborations. Please feel free to connect with me.

Email / GitHub / LinkedIn / Google Scholar

Research Interests

My research interests lie at the intersection of Multimodal AI, Computer Vision, and Edge Computing, with a focus on:

Multimodal Fusion: Integrating vision, language, and spatial inputs for robust AI systems.
Self-Supervised Learning: Developing methods for learning representations with minimal supervision.
Model Distillation: Designing lightweight models for edge deployment.
Interactive AI: Building prompt-driven interfaces for human-AI collaboration.
Uncertainty Quantification: Enhancing model robustness through uncertainty-aware learning.

Education

Ecole Nationale Supérieure de Techniques Avancées (ENSTA Paris)(QS=38)
Master of Engineering in Information and Communication Sciences & Technologies

Sep 2019 - Dec 2022
Chang'an University, Xi'an, China (Member of Project 211
Bachelor of Automation (Information and Control)

Sep 2016 - Sep 2019

Research Experience

Interactive Prompt-Based Object Detection with Cross-Modal Fusion
Developed a cross-modal framework integrating user prompts (e.g., clicks) with Vision Transformers via cross-attention and contrastive learning, enhancing adaptability for industrial inspection tasks.

Jan 2024 - Present
CNN Model Distillation Based on SAM for Industrial Anomaly Detection
Designed a lightweight CNN by distilling knowledge from Segment Anything Model (SAM), achieving 91.3% F1-Score on a high-resolution PCBA dataset with 50ms inference latency on edge devices.

Jun 2022 - Dec 2023
Semantic Image Synthesis with CGANs & Transformers
Achieved SOTA on Cityscapes, COCO-Stuff, and ADE20K via frequency-aware loss. Submitted to CVPR 2023.

May 2022 - Oct 2022

Research Experience

Interactive Prompt-Based Object Detection with Cross-Modal Fusion
Lead Researcher
Developed a cross-modal framework integrating user prompts with Vision Transformers via cross-attention and contrastive learning. Leveraged shared features for target classification and rotation prediction, ensuring real-time performance. Enhanced adaptability for industrial inspection tasks.

Jan 2024 - Present
Adaptive CNN Distillation from SAM for Industrial Anomaly Detection
Lead Researcher
Designed a lightweight CNN by distilling knowledge from the Segment Anything Model (SAM) for high-resolution industrial anomaly detection. Innovatively introduced a threshold-adaptive network structure to eliminate manual parameter tuning, enhancing system stability and robustness.

Jun 2022 - Dec 2023
Semantic Image Synthesis with CGANs & Transformers
Research Intern
Achieved SOTA on Cityscapes, COCO-Stuff, and ADE20K via frequency-aware loss. Submitted to CVPR 2023.

May 2022 - Oct 2022
Neural Architecture Search (NAS) for Deep Morphological Networks
Research Intern
Developed pseudo-morphological layers using AutoML to enhance CNN performance in edge detection and segmentation (BSD500 dataset). Published in Pattern Recognition (SCI Q1, IF=8.518).

May 2020 - Jul 2022
Self-Supervised Learning for Semantic Segmentation
Research Assistant
Developed a self-supervised pipeline using contrastive learning on COCO, achieving state-of-the-art (SOTA) segmentation performance on Cityscapes and BDD100k with minimal supervision. Submitted to Pattern Recognition (SCI Q1).

Apr 2021 - Jan 2022
Robust Semantic Segmentation via Superpixel-Mix
Research Intern
Proposed Superpixel-mix, a data augmentation method reducing model uncertainty by 18% under distribution shifts (e.g., adverse weather). Accepted at BMVC 2021 (CCF-B).

Jan 2021 - Aug 2021

Publications

Learning Deep Morphological Networks with Neural Architecture Search
Yufei Hu, N. Belkhir, J. Angulo, A. Yao, G. Franchi. Pattern Recognition, 131, 108893, 2022 (SCI Q1, IF=8.518).

Paper Link Code Link
Robust Semantic Segmentation with Superpixel-Mix
G. Franchi, N. Belkhir, M. L. Ha, Yufei Hu, et al. BMVC 2021, 2021 (CCF-B).

Paper Link

Patents

Real-time Industrial Foreign Object Detection via Vision Foundation Models
Application Number: CN202410901829.2

Developed a vision foundation model-based framework for real-time detection of foreign objects in industrial environments, achieving high precision in edge deployment.
Height Data Fusion with Texture-Guided Information Recovery and Real-Time Super-Resolution
Application Number: CN202410443098.1

Innovated a multi-sensor fusion method to reconstruct high-resolution height maps using texture guidance, enhancing accuracy and efficiency in 3D industrial inspection.
Texture-Guided Multi-Sensor Kalman Fusion for 3D Object Reconstruction
Application Number: CN202410442633.1

Designed a fusion framework integrating RGB, depth sensors, and gradient information, guided by RGB to enable high-precision 3D height data fusion using a Kalman filter strategy, providing robust support for 3D industrial inspection.
Height Data Restoration via Gradient-Integrated Fusion
Application Number: CN202410443406.0

Proposed a gradient-aware restoration algorithm for height data recovery in low-resolution industrial scans.