基于多属性分析的语义分割研究

The relevant theories and methods of semantic segmentation are of great value in academic research and industrial application. In recent years, some progress has been made in semantic segmentation, but some issues for semantic understanding still exist: the issue of efficient representation of semantic objects, accurate analysis of differences for semantic attributes, and collaborative decision making with multiple semantic attributes. Aiming at the issues above, this project puts forward difference analysis on semantic attributes for semantic segmentation from both the theory and method. The main innovations include: 1) a multi-level context-sensitive model and a semantic edge detection model will be constructed to model the local and global context relationships and the structural relationships of semantic objects; 2) an adaptive depth convolutional neural network model will be proposed to enhance the performance of difficult samples and improve the accuracy of semantic object based on difference analysis on semantic attributes. 3) on the basis of structural context analysis and decision making, semantic discriminant information and edge information are jointly learned under the unified model. The previous works and initial experimental results have demonstrated the feasibility and effectiveness of the techniques in the proposed project.

语义分割研究在学术和工业上均具有重要研究价值。近年来，相关研究在语义信息挖掘层面仍然存在如下问题：语义对象高效表征、语义属性差异准确分析和多语义属性的联合决策。针对上述问题，本项目提出基于多属性分析的语义分割方法，主要创新点包括：1）联合局部、全局和语义对象的结构关系进行建模，构建高效多层级上下文考量模型和语义边缘检测模型；2）提出自适应深度的卷积神经网络模型并对困难样本进行增强学习，从属性差异分析出发提高语义对象检测精度；3）基于上下文分析决策模型，对语义判别和边缘信息进行统一表达和联合学习，实现语义分割。工作基础和初步实验结果均表明了该项目所提出方案的可行性和有效性。

场景解析是计算机视觉领域的基础研究问题，近年来，面向国防、军事、安全等临地安防领域快速发展亟需，相关理论和技术呈现出了新的问题和难点。基于此，本项目围绕语义对象高效表征、语义属性差异准确分析和多语义属性的联合决策等难题展开研究，具体包括:1）在多模态语义特征高效表征方面，开展了特征选择-融合策略下的多模态语义特征表征和面向非对齐分辨率的多模态数据表征研究。2）在场景目标语义建模方面，开展了基于元学习特征聚合的语义建模和提高变化检测语义表征能力算法研究。3）在跨域语义特征关联学习方面，开展了基于注意力图卷积网络的语义特征关联学习和基于原型生成网络的自适应开放域识别方法研究。4）在少标注样本下的语义分割方面，开展了基于自监督学习的光谱特征表达和基于原型队列学习的遥感小样本语义分割框架研究。本项目在场景解析理论、学术成果、学术交流和人才培养方面均取得了显著成效。