JWJiesheng Wu

研究Research

研究方向Research directions

围绕视觉感知、多模态融合与语义理解，研究复杂真实环境中的目标检测、分割和跨模态学习问题。My work studies object detection, segmentation, and cross-modal learning in complex real-world environments, spanning visual perception, multimodal fusion, and semantic understanding.

01 · 视觉感知01 · Visual perception

显著性与伪装目标理解Salient and camouflaged object understanding

研究模型如何在显著目标与背景高度混淆、外观变化显著或场景信息不完整的情况下识别并分割目标。关注弱监督、无监督、检索增强与生成模型等方法，以及轻量化和高分辨率视觉模型。

I study how models recognize and segment objects when foreground and background are highly confusing, appearances vary significantly, or scene information is incomplete. Current interests include weakly supervised and unsupervised learning, retrieval augmentation, generative models, lightweight architectures, and high-resolution vision.

  • RGB、RGB-D、RGB-T 显著性目标检测/语义分割等RGB, RGB-D, and RGB-T salient object detection
  • 伪装与协同伪装目标检测Camouflaged and co-camouflaged object detection
  • 水下、遥感与高分辨率视觉Underwater, remote-sensing, and high-resolution vision
  • 二值图像分割（如表面缺陷检测、异常检测、玻璃/镜子分割）Dichotomous segmentation and surface defect detection

02 · 多模态学习02 · Multimodal learning

跨模态融合与大模型Cross-modal fusion and foundation models

探索 RGB、深度、热红外和语言等异构信息之间的对齐、交互与互补机制，并研究多模态大语言模型在视觉任务中的适配与泛化。

I explore alignment, interaction, and complementary representations across RGB, depth, thermal, and language modalities, as well as adaptation and generalization of multimodal large language models for visual tasks.

  • RGB-T / RGB-D-T 图像分割RGB-T and RGB-D-T image segmentation
  • 视觉语言表示与对比学习Vision-language representation and contrastive learning
  • 多模态大语言模型用于视觉任务Multimodal LLMs for visual tasks
  • 跨模态检索与生成式感知Cross-modal retrieval and generative perception

03 · 交叉应用03 · Interdisciplinary applications

医学影像与社会媒体理解Medical imaging and social-media understanding

将分割、检测用于医学影像和将视觉语言建模方法用于社会媒体场景，包括息肉、超声、PET-CT 肺部肿瘤分割，以及情感分析、谣言检测和讽刺识别。

I apply segmentation, detection, and vision-language models to medical imaging and social media, including polyp, ultrasound, and PET-CT lung tumor segmentation, alongside sentiment analysis, rumor detection, and sarcasm recognition.

  • 息肉与消化道图像分割Polyp and gastrointestinal image segmentation
  • 超声图像分割Ultrasound image segmentation
  • PET-CT 肺部肿瘤分割PET-CT lung tumor segmentation
  • 脑疾病预测Brain Network
  • 情感、谣言与讽刺识别Sentiment, rumor, and sarcasm recognition

研究理念Research principles

面向真实问题，重视可复现性Real problems, reproducible research

课题组鼓励从明确问题出发，兼顾方法创新、工程实现与严谨实验。高质量研究应当能够被理解、验证并复用，代码实践与论文写作同样重要。

Our group starts from well-defined problems and balances methodological novelty, engineering quality, and rigorous experimentation. Strong research should be understandable, verifiable, and reusable; implementation and writing matter equally.