融合Kernel PCA形状先验信息的变分图像分割模型

Variational image segmentation incorporating Kernel PCA-based shape priors

目的基于能量最小化的变分图像分割方法已经受到研究人员的广泛重视,取得了丰硕成果。但是,针对图像中存在的噪音污染、目标被遮挡等情况,则难以正确分割。引入先验形状信息是解决该问题的一个重要方向,但是随之而带来的姿态变化问题是一个难点。传统的做法是在每步迭代过程中单独计算姿态变换参数,导致计算量大。方法在基于Kernel PCA(KPCA)的形状先验模型基础上,提出一种具有内在的姿态不变性的KPCA形状先验模型,并将之融合到C-V变分图像分割模型中。结果提出模型无须在每步迭代中显式地单独计算姿态变换参数,相对于C-V模型分割正确率能够提高7.47%。同时,针对KPCA模型中计算高斯核函数的参数σ取值问题,也给出一种自适应的计算方法。结论理论分析及实验表明该模型能较好地解决先验形状与目标间存在的仿射变化问题,以及噪音、目标被遮挡等问题。

Objective Variational image segmentation methods, which are based on energy minimization process, have received significant attentions for years and gained fruitful achievements. However, the use of image information alone often leads to poor segmentation and results in presence of noise, clutter, or occlusion. Introducing shape prior to contour evolu- tion process has been shown as an effective way to address these problems. However, problems associated with this method is nontrivial. The traditional solution is to estimate several pose parameters within each step of level set iteration. This solution is complicated and time consuming. Method Based on the kernel principal component analysis （KPCA） shape model, we propose a novel KPCA-based shape prior model with intrinsic pose invariance, and we then combine it with C-V image segmentation model. Result The complete segmentation model explicitly eliminates pose parameter estimation during level set iteration. Furthermore, segmenting correct ratio is increased by 7.47% compared with C-V model. We present an adap- tive method to calculate parameter ofor the Gaussian kernel in KPCA shape model. Conclusion Experimental results show the robustness of the combined model against noise, clutter, or occlusion and the ability to deal with the affine pose variance between prior shapes and object to be detected.