Licheng Jiao

Research Overview

Licheng Jiao focuses on intelligent perception in remote sensing, evolutionary optimization, and brain-inspired computing. His recent work spans from foundational neural architecture search and meta-learning to applied visual grounding and synthetic aperture radar analysis. Currently, he investigates multimodal scene understanding and adaptive optimization frameworks for complex, real-world environments.

Research Trajectory

The recent trajectory of Licheng Jiao’s research reflects a sustained focus on adapting advanced machine learning paradigms—ranging from vision transformers to evolutionary algorithms—for specialized and challenging domains. A significant portion of his work centers on remote sensing and synthetic aperture radar (SAR) image analysis. In 2025, he developed RegionMatch: Pixel-Region Collaboration for Semi-Supervised Semantic Segmentation in Remote Sensing Images, a framework that enforces consistency between individual pixel predictions and broader object-level features to generate high-quality pseudo-labels from limited annotated data. Addressing the complexities of multi-modal satellite imagery, he introduced Multi-Expert Learning Framework with the State Space Model for Optical and SAR Image Registration. This approach leverages the efficiency of State Space Models alongside a multi-expert strategy to align optical photos with radar images in texture-poor landscapes, contributing to the broader area of Mitigating Speckle Noise in Synthetic Aperture Radar.

Further streamlining SAR analysis, he proposed AutoPolCNN: A neural architecture search method of convolutional neural network for PolSAR image classification, which automates the design of classification models for polarimetric SAR data. Expanding his remote sensing research into vision-language tasks, his group designed CLIP-Based Grid Features and Masking for Remote Sensing Image Captioning. By utilizing visual grid features and a random masking strategy, this transformer-based model improves the extraction of relational context between words and image regions, intersecting with research directions such as Language Guided Image Restoration and Enhancement.

Parallel to his work in remote sensing, Jiao has advanced the theoretical and applied foundations of optimization and neuromorphic computing. In 2025, he formulated A two-stage strategy for brain-inspired unsupervised learning in spiking neural networks, creating a multi-stage learning mechanism that improves the flexibility and adaptability of low-power Spiking Neural Networks (SNNs). In the signal processing domain, he addressed cross-domain distribution shifts with PSRNet: Few-Shot Automatic Modulation Classification Under Potential Domain Differences, improving the reliability of modulation classification across varying radio frequency devices. His algorithmic research also heavily emphasizes meta-learning for optimization. He presented Learning Evolution via Optimization Knowledge Adaptation, introducing the OKAEM framework. This model replaces hard-coded evolutionary heuristics with learnable, attention-based neural modules that self-tune in real-time on new tasks.

Jiao’s expertise extends into complex 3D vision, pose estimation, and egocentric scene understanding. To handle real-world occlusions and missing viewpoints, he designed DeProPose: Deficiency-Proof 3D Human Pose Estimation via Adaptive Multi-View Fusion, a flexible architecture that avoids the cumulative errors typical of multi-stage pose estimation networks. He also constructed EgoSplat: Open-Vocabulary Egocentric Scene Understanding with Language Embedded 3D Gaussian Splatting. By combining temporal tracking with robust feature aggregation, EgoSplat effectively filters out transient objects to maintain semantic consistency in first-person videos.

Moving into 2026, Jiao’s collaborative work has continued to push the boundaries of efficiency and cross-modal reasoning. He participated in developing Foreground-Aware Token Routing Vision Transformer for Real-Time Satellite Video Tracking, which utilizes a shrink-and-expand token routing strategy to focus computational resources on foreground targets, achieving high frame rates without sacrificing accuracy. This aligns closely with the research area of Precise Localization for Arbitrarily Oriented Objects. In the realm of 3D vision-language integration, he contributed to DSGCR: Decomposed Spectral Geometry-Aware Cross-Modal Semantic Representation for 3D Visual Grounding. This model explicitly separates geometric relations into symmetric and antisymmetric components to capture high-frequency spatial details, advancing the domain of Anisotropic Multiscale Representations for Geometric Features. Finally, building on his prior optimization research, he co-developed Task-free Adaptive Meta Black-box Optimization, an online meta-optimizer that transforms evolutionary operators into differentiable modules, eliminating the need for expensive pre-training on massive task distributions.

Current Directions

Recent work investigates efficient architectures for real-time visual tracking, cross-modal representations for 3D scene understanding, and adaptive meta-optimization. The most recent papers in the evidence address the optimization of Vision Transformers through dynamic token routing strategies for satellite video analysis, prioritizing computational efficiency in resource-constrained environments. Additionally, ongoing research explores the use of decomposed spectral geometry to better align visual and linguistic features for precise 3D object grounding. In the algorithmic domain, current efforts focus on task-free meta-learning frameworks that adaptively update evolutionary optimization operators online, allowing black-box optimizers to learn and adjust their strategies during the optimization process itself without relying on extensive prior task distributions.