Research
Our lab focuses on developing cutting-edge DL algorithms and artificial intelligence systems that solve real-world problems across multiple domains.
Current Research Directions
Our active research spans multimodal perception, efficient and hardware-aware AI, and learning-based low-level vision. A number of these efforts are conducted with industry partners or are currently under peer review, so detailed results and publications will be released as the work matures.
Vision–Language Models
Multimodal vision–language models (VLMs) for joint image–text understanding, visual grounding, and reasoning, including the adaptation and alignment of large VLMs for domain-specific perception.
Object Detection & Segmentation
Robust object detection and semantic/instance segmentation, with an emphasis on accuracy and reliability under challenging, real-world visual conditions.
Efficient AI on Edge Devices
Lightweight neural architectures together with model compression and quantization for low-latency, energy-efficient inference on resource-constrained edge platforms.
Hardware–Software Co-Design
Designing hardware-friendly AI models from a deployment-aware perspective, jointly optimizing algorithms and accelerators for SoC/FPGA targets under memory, throughput, and power budgets.
RAW-Domain Learning & AI-ISP
AI-based image signal processing that learns directly in the RAW sensor domain, rather than the conventional sRGB domain, to fully exploit sensor information for downstream vision tasks.
Foundation Vision Models for Low-Level Vision
Adapting large-scale foundation vision models as priors and backbones for downstream low-level vision applications such as restoration, enhancement, and reconstruction.
Implicit Neural Color Representations
Implicit neural representation (INR)-based 3D look-up tables (3D LUTs) for learnable color transformation, color-space conversion, and color-difference / chromatic-aberration correction.
Cross-Camera Auto-Annotation for Autonomous Driving
Automatic annotation of multi-camera, surround-view autonomous-driving imagery, exploiting cross-camera geometric and temporal consistency to generate high-quality perception labels while greatly reducing manual labeling effort.
Research Focus Areas
Hyperspectral Image Enhancement
Developing advanced algorithms to enhance the quality and information content of hyperspectral imagery by fusing low-resolution hyperspectral data with high-resolution multispectral images, generating high-resolution outputs with improved spatial and spectral details for more accurate remote sensing analysis.
Image Deraining
Creating innovative techniques for removing rain streaks and raindrops from images and videos, improving visibility and enabling better performance of computer vision systems in adverse weather conditions.
Image Dehazing
Developing specialized algorithms to remove haze interference from images, restoring clarity and color fidelity while preserving structural details for improved visibility in applications ranging from photography to autonomous vehicle navigation.
Super Resolution
Developing advanced algorithms to generate high-resolution images from low-resolution inputs, recovering fine details and enhancing visual quality beyond the limitations of capture devices.