融合超像素与稀疏约束的水平集图像分割模型研究

Image segmentation can be described as using low-level visual features and extracting region of interest of user from the complex background environment, which is the advanced expression of the original pixels to semantic objects and a basic problem in computer vision. In recent years, the level set segmentation method has made great progress in image segmentation by integrating shape prior constraints and local area information to construct energy functional. However, the current level set model of image segmentation is susceptible to noise and lacks robustness; the precision of object boundary extraction is also needed to improve; the convergence is slow and the computation is time-consuming. Inspired by the theory of sparse representation, based on the superpixel technology, some researches about level set based image segmentation model will be conducted, and the contents are as follows: Firstly, based on the classical superpixel segmentation algorithm, a function of distributive entropy is defined to improve the precision of object. Secondly, a feature basis function is defined for every superpixel, and the linear combination of basis functions is used to replace the traditional heaviside function. Finally, a new energy function for superpixel is constructed, and the level set based image segmentation model is proposed by integrating superpixel and sparsity constraint. The above researches could offer a new perspective and method for image segmentation, and can enrich the theory of image segmentation and image analysis. Furthermore, they could also improve the robustness and precision in segmenting objects.

图像分割是利用低层视觉特征，从复杂的背景环境中提取用户感兴趣的区域，是原始像素对语义对象的高级表达，也是计算机视觉中的一个基本问题。近年来，水平集分割方法通过整合形状先验约束和局部区域信息构建能量泛函实现图像分割，取得了极大进展。然而，目前的水平集模型易受噪声影响，演化不够稳健；目标边界分割精确度仍有待提高；收敛速度慢，算法耗时。受稀疏表示理论启发，基于超像素技术，本项目对水平集图像分割模型进行研究，其内容如下：①基于经典的超像素分割技术，定义分布熵函数，提升超像素在边界保持方面的表现。②对每个超像素定义一个特征基函数，用基函数的线性组合代替传统的Heaviside函数。③构建超像素水平集能量泛函，提出基于超像素和稀疏约束的水平集图像分割模型。该研究能为图像分割提供一个新的视角和方法，也能丰富图像分割和分析理论，提高目标分割的精确性和稳健性。

水平集分割模型由于有成熟的理论基础和良好的可解释性，对该模型研究仍具有重要意义。针对当前水平集分割模型的分割精度不够理想、易受噪声和非目标物体干扰等问题，基于超像素和稀疏表示方法，本项目对水平集图像分割模型展开研究，其主要内容如下：①超像素分割优化框架研究。基于经典的超像素分割方法，通过定义信息度量函数提出了超像素边界优化框架。②稀疏性特征应用到经典活动轮廓模型研究，分别把稀疏表示方法应用到经典参数活动轮廓模型和水平集分割模型中。③基于超像素分割方法，分别提出了基于局部相似因子和显著性映射的超像素活动轮廓模型、融合超像素和像素的联合约束图像分割模型、超像素/像素协同约束和稀疏分解的活动轮廓模型。④融合卷积神经网络和活动轮廓模型研究及其在医学影像目标分割应用。.基于上述研究内容，提出的基于信息度量函数的超像素边界优化框架，在超像素块数目较小的情况下对经典方法的提升更显著，可以弥补经典算法的目标缺失缺点，为后续的图像分割提供了良好基础。利用稀疏表示方法，分别融合到参数和水平集分割模型中，提升了演化的稳健性，达到减小噪声和非目标物体的影响。利用超像素方法，既能快速收敛到目标物体附近，又对噪声稳健。和卷积神经网络进行有效融合，能有效处理分割结果的非光滑性，且能灵活的利用先验信息对分割结果进行进一步的细化。该研究能为图像分割方法提供一个新的参考，也能丰富图像分割和分析理论，提高目标分割的精确度和稳健性。