基于非监督决策树的模糊图割模型的彩色图像分割研究

Unsupervised color image segmentation is widely applied in the machine vision of industry, agriculture, medicine fields, and it is also an important method for intelligent production. Although the traditional methods working with the model of the hierarchical graph cut can meet various segmentation requirements in different scenes of fields, they still has some problems like NP-hard, high complexity, heavy computation, rigid segmentation and so on, due to ignore the effective structure of model and the fuzzy characteristics of the images. Combining hierarchical classification model of the decision tree with the theory of fuzzy segmentation, this project can solve those problems. First of all, we analyze the impact between objects, foundation, convergence and segmentation results, then we develop unsupervised decision tree model of fuzzy graph cut for color image segmentation; Secondly, we utilize the fuzzy theory to guide the extraction and fusion process of color and texture feature, as well as the implementation of fuzzy graph cut; Finally, we build the parameter optimization mechanism of the model, and realize the coordinated optimization. The research results will be tested by plenty of simulated images and real images. In this project, we will develop novel efficient algorithms for solving above-mentioned segmentation problems, and contribute to new concept, theory and methods for unsupervised color image segmentation. In addition, this project will promote the development of unsupervised graph cut, fuzzy segmentation and hierarchical segmentation.

非监督彩色图像分割被广泛应用于工业、农业、医学等领域的机器视觉中，是智能化生产的重要手段。传统的层次化图割模型虽能满足不同领域场景的多样化分割需求，但未考虑高效的模型结构和图像的模糊特性，存在多标签求解NP难，复杂度高，计算量大，硬划分等缺点。本项目将决策树的层次化分类模型与模糊分割理论相结合来解决上述问题。首先，分析分割对象、分割依据、收敛策略对层次化图割算法的效率和结果影响，构建针对彩色图像分割的非监督决策树的模糊图割模型；其次，利用模糊理论指导颜色和纹理特征的提取和融合以及模糊图割的设计和实现，建立基于局部和整体模型的参数优化机制，实现各模块参数之间的协调匹配；最后在仿真图像和真实图像上进行了大量测试。项目拟研究解决层次化图割问题的高效方法，为非监督彩色图像分割提供新概念、新原理和新方法，同时也将推动非监督图割，模糊分割和层次化分割等方面的自身发展。

彩色图像分割作为图像处理和分析的关键步骤，被广泛应用于工业、农业、医学、军事、交通等领域的机器视觉中。然而，非监督彩色图像分割技术却面临着模糊目标难以精确分割，硬划分，多标签求解NP难，计算量大的问题。结合层次化的决策树分类模型和模糊熵分割理论，能有助于解决非监督彩色图像分割时存在的问题。为此，本项目提出了非监督层次化模糊熵图割的彩色图像分割方法。本项目具体研究内容和创新点包括：解析图像盲复原的最小优化问题，提出了贝叶斯迭代联合双边滤波的图像复原方法，实现待分割图像的预处理；提出两标签分割的快速递推模糊2-划分熵图割算法，用模糊2-划分熵时的隶属度函数来设置图割的似然能，使基于阈值的分割方法与基于空间相关性的分割方法优势互补，提高分割精度，期间还提出了递推策略加速最大模糊2-划分熵的计算，从而解决了两标签分割的硬划分和计算量大的问题；将两标签的模糊熵2-划分图割算子与自顶向下的非监督层次化的决策树分割策略相结合，通过选择恰当的通道和划分区域，迭代地对彩色图像实施2-划分，直到算法收敛，划分数自动确定，从而解决了非监督彩色图像分割多标签求解NP难，硬划分，计算量大的问题。本项目的研究成果不但能应用于彩色图像分割，还能用于工程实践，为非监督彩色图像分割提供新概念，新原理和新方法，同时也推动了非监督图割，模糊分割和层次化分割等方面的自身发展。