低分辨率图像语义分割问题是制约实际应用中分割精度的瓶颈之一。本项目拟从图像信息重构角度，研究低分辨率图像语义分割问题。首先，拟构建任务递归式的学习框架，探究低分辨率图像分割与超分辨任务间的相关性；拟设计基于递归网络的结果迭代优化模型，并通过多模态数据整合单元强化跨任务的特征传递。其次，拟提出基于嵌入空间特征重构的低分辨率分割算法。拟设计特征嵌入网络，实现低分辨率图像的特征嵌入提取与重构，并通过建模图像场景语义信息，对嵌入特征进行有效筛选；拟探索对抗训练与蒸馏学习相结合的训练策略，实现嵌入网络的高效学习。最后，拟提出基于自适应超分辨的小目标区域分割算法，提升关键小目标区域的分割精度；拟研究全局与局部模型相结合的算法框架，探索高效的跨层级特征共享与结果融合机制；拟研究全监督学习与深度强化学习方法，实现自动的小目标检测、超分辨与分割。拟通过上述研究，为低分辨率图像语义分割问题提供理论与实践支撑。

The challenges delivered by low-resolution images have significantly limited the semantic segmentation accuracy in real-world applications. This project proposes to study low-resolution image semantic segmentation through image information reconstruction. First, a task-recursive learning framework is proposed to explore the correlation between low-resolution image segmentation and super-resolution. An iterative refinement model will be designed using recurrent networks with a multi-modal data integration unit to ensure cross-task feature transformation. Second, an information reconstruction algorithm in the feature embedding space will be developed to facilitate more accurate low-resolution image segmentation. A feature embedding network will be designed to perform feature extraction and reconstruction for low-resolution images, where the reconstructed image features will be further selected based on the encoded semantic representation of the input images. The collaboration between adversarial training and distillation learning will also be explored for efficient network optimization. Finally, this project will present an adaptive super-resolution algorithm to further improve segmentation accuracy on key and small object regions. To this end, the integration between global and local segmentation models will be studied to enforce cross-level feature sharing and result aggregation mechanisms. Besides, both fully supervised and deep reinforcement learning strategies will be leveraged to enable automatic small object detection, super-resolution, and segmentation. The key findings and algorithm designs of the above proposals will provide a theoretical basis and practical reference for low-resolution semantic segmentation.