吴永伟

吴永伟

深圳技术大学 集成电路与光电芯片学院，助理教授

College of Integrated Circuits and Optoelectronic Chips, Shenzhen Technology University，

Assistant Professor

吴永伟，博士，深圳技术大学助理教授，电子元器件高级工程师，深圳市高层次人才。博士毕业后于华星光电、北京大学从事新型显示技术研究，聚焦4K/8K LCD、Mini/Micro-LED、量子点等产品与新型显示技术开发。截至目前，已授权中国/美国发明专利 41 项（其中第一作者授权 33 项）；发表科研论文超 20 篇；主持参与国家自然科学基金-青年基金、中国博士后科学基金、国家重点研发计划等科研项目与企业研发项目多项。

Yongwei Wu, Ph.D., is an Assistant Professor at Shenzhen Technology University and a Senior Engineer of Electronic Components. He is recognized as a Shenzhen High-Level Talent. Following his doctoral studies, he engaged in research on novel display technologies at TCL China Star Optoelectronics Technology (CSOT) and Peking University. His research interests focus on the development of 4K/8K LCD, Mini/Micro-LED, and Quantum Dot technologies. To date, he has been granted 41 invention patents in China and the United States, including 33 as first inventor, and has published more than 20 papers. He has also led and participated in multiple research projects, including the National Natural Science Foundation of China (Youth Fund), the China Postdoctoral Science Foundation, the National Key R&D Program of China, as well as several industry R&D projects.

题目

墨滴到像素：数据高效的色转换层打印工艺优化与实时 FPGA 图像增强

摘要：

人工智能正在重塑显示技术，但在实际部署中面临材料研发数据匮乏与边缘端硬 件资源受限的双重挑战。本报告展示一种面向显示链路的轻量化 AI 策略，覆盖 喷墨打印色转换层(CCL)的工艺优化与嵌入式图像增强。针对小样本实验，我们 以随机森林建模并排名关键配方与工艺参数，结合贝叶斯优化进行虚拟筛选，获 得矩阵含量 55%到 65%与膜厚 10 到 14 μm 的优选窗口，参数寻优周期缩短约 60%(R²=0.58)。同时提出面向 FPGA 的轻量低照度增强 CNN，通过通道剪枝与 INT8 量化将参数降至 1/64，并用线性查表近似激活；在保持熵 7.01 的同时将 LOE 由 79.77 降至 66(提升 17%)，实现全高清 120 Hz、1.4 W 实时处理。

Title

From Inkjet Droplets to Pixels: Data-Efficient CCL Printing Optimization and Real-Time FPGA Image Enhancement

Abstract：

Artificial Intelligence is reshaping the display industry, yet its practical implementation faces challenges ranging from data scarcity in material research to hardware constraints in edge processing. This talk presents an AI-enabled, data-efficient workflow spanning display manufacturing and real-time imaging. For inkjet-printed color conversion layers (CCLs), we address a high-dimensional process space with fewer than 50 experiments. A Random Forest surrogate, validated by leave-one-out cross validation (R² = 0.58), ranks key formulation and process factors. Using a composite quality factor that jointly reflects color gamut, efficiency, and spectral stability, Bayesian optimization performs virtual screening and identifies a narrow optimum window, with matrix content of 55% to 65% and thickness of 10 to 14 μm, cutting optimization time by about 60%. For deployment-oriented imaging, we introduce a lightweight low-light enhancement CNN tailored for FPGA. With channel compression (32 to 4), pruning, INT8 quantization, and linear LUT activation approximation, the model maintains entropy at 7.01 (baseline 7.16) while improving illumination fidelity by 17% (LOE 79.77 to 66). An FPGA prototype achieves real-time 1080p enhancement at 120 fps with 1.4 W power and modest on-chip resource usage.