李强

西安电子科技大学，助理教授

Xidian University, Assistant Professor

李强，西安电子科技大学助理研究员，中国光学工程学会光显示专业委员会委员，中国图象图形学会三维成像与显示专业委员会委员。一直从事光场显示、AR显示、智能图像处理等相关理论和技术的研究工作，入选2023年首批“国家资助博士后研究人员计划”，两年来主持国家级、省级等项目10项，包括国家自然科学基金青年科学基金项目、陕西省自然科学基础研究计划资助项目、陕西省秦创原引用高层次创新创业人才项目、小米揭榜挂帅-2025科研专项等。在国际权威期刊和国际专业会议上发表SCI/EI学术论文30余篇；申请中国发明专利20余件，其中授权专利10件；担任Photonics的 Guest Editor，担任Elsevier、Optica、IEEE、Wiley旗下等十多个国际顶级/权威期刊的审稿人。曾获得全国大学生光电设计竞赛全国优秀指导教师、中国国际大学生创新大赛省级优秀指导教师、挑战杯中国大学生创业计划竞赛省级优秀指导教师、大学生创新创业项目国家级优秀负责人、四川省优秀毕业生、四川大学优秀博士论文、四川大学第十四届十佳学术之星、博士研究生国家奖学金等六十余项荣誉。

Qiang Li is an assistant professor at Xidian University. He received his PhD degree from Sichuan University in 2023. He has published more than 30 papers on information displays. He serves as a Professional Committee of the Chinese Society for Optical Engineering and a member of the Three-Dimensional Imaging and Display Professional Committee of the China Society of Image and Graphics. His research focuses on the theories and technologies related to light field displays, AR displays, and intelligent image processing. He has led ten national and provincial research projects, such as the National Natural Science Foundation of China, the Natural Science Basic Research Program of Shaanxi, and the Xiaomi - 2025 Research Special Project. Dr. Li has served as a Guest Editor for the Photonics and as a reviewer for over ten international journals under publishers including Elsevier, Optica, IEEE, and Wiley. He has received more than sixty honors and awards.

题目

光场智能三维处理与显示

摘要：

人工智能与光学成像的深度融合正推动三维视觉技术迈向高精度、高效率与高质量新阶段，在此背景下，本次报告分享我们课题组近期在光场智能处理与显示方向的三项研究进展。首先，提出基于Mamba的轻量级显著性检测网络LFSamba，引入显著性引导的序列扫描机制，在保持线性计算复杂度的同时，相较CNN与Transformer方法提升精度超60%、参数量减少50%；其次，面向光场高帧率需求，设计Triple I-3D Net插帧网络，通过自适应感受野与统一优化策略，高质量合成中间微图像阵列，显著提升动态显示性能；最后，针对透镜阵列旋转误差引起的体素扩散，构建基于同名像素-体素分布学习的生成对抗网络PVGAN，首次实现多误差耦合机制的定量建模与端到端校正，无需迭代即可批量重建三维内容，其空间分辨率较未校正系统提升3–4倍，显著改善显示质量。这些工作为光场医学可视化、沉浸式三维显示等应用提供关键技术支撑。

Title

Intelligent three-dimensional processing and display of light field

Abstract：

The deep integration of artificial intelligence and optical imaging is driving 3D vision technologies toward unprecedented levels of accuracy, efficiency, and visual fidelity. In this context, this talk presents three recent advances from our group in intelligent light field processing and display. First, we propose LFSamba, a lightweight Mamba-based network for 3D salient object detection in light fields. By introducing a saliency-guided sequential scanning mechanism, LFSamba leverages Mamba’s linear-complexity global modeling capability to achieve over 60% higher detection accuracy while reducing model parameters by 50% compared to CNN- and Transformer-based methods. Second, to address the demand for high frame rates in light field displays, we design Triple I-3D Net, an interpolation network that synthesizes high-quality intermediate elemental image arrays through adaptive receptive fields and a unified motion-synthesis optimization strategy, significantly enhancing dynamic 3D display performance. Third, to mitigate voxel diffusion caused by rotational misalignments of the lens array, we develop PVGAN, a generative adversarial network based on homologous pixel–voxel distribution learning. PVGAN achieves, for the first time, quantitative modeling and end-to-end correction of coupled multi-error mechanisms. Notably, it enables batch-wise 3D reconstruction without iterative preprocessing, reaching a three to four times improvement in spatial resolution over the uncorrected system and substantially enhancing display quality in terms of sharpness. These works provide key technical support for applications such as light field medical visualization and immersive 3D display.