Inclusion Degree And Kernel Descriptor Based Active Contour Models

Image segmentation is one of the most important parts in image processing. It has been widely used in plenty fields such as pattern recognition, object tracking and detection, image restoration and feature extraction. In recent years, active contour models become one of the most popular tools to solve image segmentation task. The main works are as follows:1. We start from some theory analysis and discussions on classical active contour algorithms and then propose two novel active contour models. Firstly, we propose an inclusion degree based nonparametric statistical active contour model(INDAC). The method connects statistical image domain with fuzzy theory. We utilize nonparametric density estimates to estimate the probability density functions of foreground and background. We formulate the overlapping rates between foreground and background through borrowing inclusion degree from fuzzy set. The two overlapping rates are used to formulate our energy function, subject to a constraint on the total length of the region boundaries. We solve the inclusion degree based optimization problem by deriving the associated gradient flow and applying curve evolution techniques. Level-set methods are used to implement the resulting evolution. The experimental results based on both synthetic and real images demonstrate that the proposed technique can solve various image segmentation problems. When compared to active contour models previously formulated to solve the same nonparametric statistical segmentation problem, it is observed that our method performs as good as the existing ones. Furthermore, the proposed method exhibits higher efficiency and consumes less evolution time.2. We introduce a kernel descriptor based interactive active contour model(KDIAC). Early active contour models are based on pixel-wised low level features which lack spatial correlations leading to bad segmentation results for heterogeneous images. High dimension features are required. However, high level features not only complicate the mathematics and the implementation of the contour evolution but also lacks efficient and meaningful similarity/dissimilarity metrics. In this paper, we introduce kernel descriptor. It incorporates spatial correlations between pixels and facilitates the computation of similarity between different features. KDIAC consists of two parts:initial contour approaching the ideal boundary is obtained through the coarse stage and we formulate the energy functional by the similarity between the pixels on the contour and the template features in the second stage. Experimental results show that our method is robust for various inputs of the users. Moreover, in some heterogeneous images, our model provides more accurate segmentation than some interactive methods such as graph-based methods.