小样本下组合性模型的结构表征与因果学习推理研究

针对组合性模型的小样本学习框架在图像和视频认知任务中存在的轮廓结构表示和推理等方面的挑战性问题，拟建立轮廓素描线计算模型，拟构建基于轮廓素描线计算模型和自监督信息的物体轮廓结构和子结构所在区域获取方法；拟提出能显式地表示出图像中物体轮廓结构和子结构的层次横向联系表示方法。针对视频中的运动物体，拟建立能描述物体运动属性的线流稀疏模型，并提出一种视频运动信息的稀疏表示方法——线流。我们拟构建当前帧和相邻视频帧中运动物体组成部分的轮廓结构所在区域获取方法；拟提出能显式地表示出视频中运动物体结构和子结构的层次横向联系表示方法。本项目拟提出因果学习推理的优化框架，该框架将学习和优化的思想引入到整个推理过程中来，有助于提升机器推理-决策的水平。在因果学习推理方面，针对复杂视觉任务，基于概率上下文相关语法模型和自底向上/自顶向下联合推理框架，拟构建具有选择注意机制和因果关系的学习推理模型和学习推理方法。

Aiming at the challenging problems of contour structural representation and reasoning in image and video cognitive tasks for the framework of compositionality model with few samples learning, a contour sketch line calculation model is proposed, and an object contour structure and substructure region acquisition method based on the contour sketch line calculation model and self-monitoring information is proposed. This project intends to propose a hierarchical horizontal connection representation method that can explicitly represent object contour structures and substructures in images. For moving objects in video, a sparse line flow model that can describe the motion attributes of the objects will be established, and a sparse representation method of video motion information is proposed, which is called line flow. We plan to build a region acquisition method for the contour structure of moving object components in the current frame and adjacent video frames. It is proposed to present a hierarchical horizontal connection representation method that can explicitly represent the contour structure and substructure of moving objects in video. This project intends to propose an optimization framework of causal learning reasoning, which introduces the ideas of learning and optimization into the whole reasoning process and helps to improve the level of machine reasoning-decision. In the aspect of causal learning reasoning, for complex visual tasks, based on probabilistic context-related grammar model and bottom-up/top-down joint reasoning framework, a learning reasoning model and learning reasoning method with selective attention mechanism and causal relationship are proposed.