Saliency-based Approaches for Automatic Image Segmentation (Diplomarbeit)


Khalil Fergani


Saliency detection has become an important technique in computer vision applications, such as region of interest segmentation, adaptive image compression and pattern recognition. To automatically detect the salient regions in images, visual attributes such as color, orientation and edges are used. Despite its potential to describe the global compositional characteristics of the image, the visual attribute texture has not been widely explored for the purpose of saliency estimation. A novel saliency approach based on the concept of statistical textural distinctiveness is proposed. Rotational invariant textural representations are extracted and used in an efficient manner to quantify the distinctiveness of each region compared to the rest of the image. The saliency of each region is then computed based on the textural distinctive- ness and a general visual constraint. The proposed method was shown to provide a strong saliency detection performance when compared to existing methods. Furthermore, to investi- gate the potential of the saliency characteristics in the vector field convolution active contour model, a new saliency-guided active contour framework is proposed. In order to improve object boundary detection in images with cluttered background, both the intensity characteristics and saliency characteristics of the image are incorporated into the vector field convolution external energy functional. In addition, an automatic saliency-based contour initialization is performed to initialize the contour close to the object of interest. Considering the saliency character- istics contributes to avoiding misconvergence to local minima and guides the active contour to converge towards the desired object boundary. The texture distinctiveness approach, as well as other saliency detection approaches, were used in the new framework. Experimental results based on a publicly available natural images data set showed that the proposed texture distinctiveness approach and the saliency-guided vector field convolution active contour pro- vide strong image segmentation performance. Furthermore, the proposed methods provided promising results when applied to magnetic resonance images (MRI) for the purpose of brain tumor segmentation.