Zihao Yang
Ph.D. Candidate @ Anhui University
Biography
Hello! I am Zihao Yang, currently a Ph.D. candidate in Electronic Science and Technology at Anhui University. My research focuses on the intersection of AI in general vision and agriculture, exploring how machines can "understand" the complex world and perform logical reasoning. My goal is to build efficient, precise, and continuously learning intelligent vision systems.
Research Interests
- General Vision Models
(Object Detection / Segmentation) - Vision-Language Models
(MLLM / Diffusion) - Visual Complex Reasoning
(Chain-of-Thought / Visual Thinking) - Visual Token Compression
(Pruning / Distillation) - Visual Continual Learning
(Memory / Fast-Slow Systems) - Embodied AI
(Robotics / Autonomous Driving)
Education
Anhui University (AHU)
Ph.D. Candidate
Sept. 2023 - Present
Electronic Science & Technology
2025 Anhui Prov. Dept. of Edu. Ph.D. Research Project Grant
Anhui Science and Tech. Univ. (AHSTU)
Bachelor
Sept. 2019 - June 2023
Optoelectronic Information Science & Engineering
Rank: 1 / 37
GPA: 3.62 / 4.0
Ranked 1st in Entrance Exam (Prelim & Re-exam)
Full Scholarship
CET-6
National Physics Competition (3rd Prize)
Outstanding Graduate
Research
Selected Projects
Remote Sensing Zero-Shot Detection & Lightweight Deployment
- Designed and implemented a deep learning-based zero-shot object detection model, utilizing cross-modal alignment technology to overcome the limitation of heavy annotation dependence for new military targets.
- Led model optimization strategies such as quantization, pruning, and knowledge distillation for domestic Ascend NPUs, implementing the Pt→ONNX→OM pipeline for efficient inference on edge computing devices.
Automotive Part Defect Detection (Dongfeng Line)
- Developed a deep learning vision model to precisely identify surface defects on automotive parts, such as scratches, cracks, deformations, and missing components.
- Integrated industrial cameras, lighting, and PLC hardware; developed software modules for image acquisition, processing, result judgment, and data visualization, automating the inspection process.
361° Shoe Sole Automatic Gluing System
- Designed a fully automated gluing system meeting 361°'s specific process requirements, integrating sole positioning, visual guidance, and precision spraying.
- Utilized binocular cameras with structured light for 3D point cloud reconstruction and instance segmentation to extract gluing areas, providing accurate paths for the robotic arm.
Provincial Natural Science Foundation — LLM-based Pest Classification
- Led the collection, cleaning, and fine-grained annotation of large-scale agricultural pest images, establishing a dataset of over 100k images across 60 categories.
- Researched and applied transfer learning and efficient fine-tuning techniques for Visual Large Models, significantly improving pest recognition with limited samples.
- Constructed a Multimodal Knowledge Graph for RAG (Retrieval-Augmented Generation) to reduce MLLM hallucinations and improve accuracy in agricultural Q&A.
National Natural Science Foundation — Wheat Disease Recognition
- Identified tiny pests and diseases in UAV aerial images of wheat using deep learning algorithms.
- Designed model optimization strategies based on domain adaptation and meta-learning, significantly enhancing generalization across different regions, climates, and planting conditions.
Lifestyle & Interests
Overwatch
Hardcore Player.
Top 500 Player
Marathon
Challenge limits, measure the city with footsteps.
Half Marathon PB
1:52:58
Active Life
Enjoying the wild and the court.
Pop Culture
Eason Chan, Jiang Wen, Wong Kar-wai.