基于随机森林和深度学习耦合模型的RGB-D图像语义标注关键技术研究

The desire to the unified optimization framework coupling feature learning and the classification model has focused the attention of the image semantic labelling community. Due to the higher complexity of the model, the larger size of parameter, and the exchange between parameter, the realization of the coupling model inevitably encounter difficulties in computational power, storage requirement, and efficient optimization. However, the low accuracy rate caused by prediction error of decision tree and the general lack of expert knowledge is a tough challenging problem for the classification method. To solve the aforementioned challenges, we investigate a host of research problems and develop a range of technical solutions in this project. First, we propose a new feature mapping method (which can preserve most of the discriminative information) to resolve the problem of feature reuse, and introduce the expert knowledge into the prediction model to improve the accuracy rate. Second, by extending the sparsity analysis method to parameter tuning, we plan to devise a new algorithm to effectively train the coupling model with a compact form. Taking the sparse representation analysis theory as background mathematical tool, we focus on the compression of the parameters for the coupling model, which facilitates to the portability improvement of the initial coupling model. Finally, in this project, we can build a complete framework for RGB-D image semantic labeling, including the coupling of random forest with feature learning and expert knowledge, compact and light-duty coupling model design. Specially, our framework can afford efficiency and high accuracy for data modeling, which will offer new theories and engineering solution for some applications of pattern recognition.

基于特征学习和多分类任务耦合的统一处理框架已成为图像语义标注领域的研究热点。在耦合模型构建过程中，受到模型复杂度高、参数维数巨大、参数间内部协作等因素的影响，不可避免地面临计算能力不足、存储空间缺乏、模型高效求解困难等问题。此外，由于随机树预测误差和专家知识缺乏等因素产生的预测低准确率是目前多分类任务处理中极具挑战的问题。为此，本项目拟针对耦合模型中的特征重用和预测准确率低问题，研究语义信息保持的特征维数映射方法和耦合专家知识的决策模型；针对耦合模型中的参数维数巨大问题，研究基于稀疏性先验知识的紧致耦合模型构建；针对模型可移植性差问题，研究基于稀疏表示模型的轻型耦合模型构建。最终形成以耦合特征学习和专家知识的优化模型、以及紧致化和轻型化耦合模型为主体算法的RGB-D图像语义标注处理框架，实现高效、高精度数据建模,为模式识别领域相关问题提供理论和技术支持。

本项目研发期间，针对所面临的关键问题展开基础理论分析和研讨，在具有共生结构特性的谱聚类算法、深度测距仪定位方法、稀疏正则化约束特征选择策略、多尺度结构特征提取、轻量级网络构建等方面，提出基于代表点的大尺度谱聚类方法，解决了传统谱聚类方法高计算复杂度问题；提出了长方体标定物的二维激光测距仪定位方法，实现测距仪的空间位置的精确计算；提出了基于线性判别分析和L_2,1范式的鲁棒特征选择方法，实现最具区别性特征的获取，并有效消除冗余特征；提出了基于热核的数据相关各向异性小波，实现多尺度内蕴结构特征的提取；提出了基于全局特征抽象的轻量级网络模型，实现主干网浅层细节特征和深层全局上下文特征的双向学习和融合，提高了图像语义分割算法的速度；提出了基于多维度特征内在联系学习模型的轻量级网络模型，实现多维度特征的内在联系学习，提高了网络模型的深度估计准确性。经过主持人与团队成员的不懈努力，已发表学术论文4篇，其中SCI论文3篇，EI论文1篇，申请国家发明专利1项；在投稿论文2篇。因此，该项目总体较顺利地完成了研究计划，达到研究目标。