Research Of Image Co-segmentation Methods Based On Active Contour Model

In recent years,image co-segmentation has gradually become a hotspot in the field of image segmentation.As a weakly supervised segmentation method,image co-segmentation methods assume that the group of images include same or similar objects.Taking full advantages of the additional information,image co-segmentation usually could achieve better effectiveness than the segmentation methods for single image.Compared to supervised segmentation methods,image co-segmentation methods has better prospects because it does not require labeling samples.Over the past ten years,image co-segmentation has greatly developed and extended in many aspects,e.g.,from Markov random field based co-segmentation methods to various theories and models based co-segmentation methods,from co-segmentation methods on pair of images to the methods on multiple images,from single-object co-segmentation to multi-object co-segmentation,from image co-segmentation to video co-segmentation.However,when the images to be segmented are complex,the effectiveness of current co-segmentation methods cold not obtain the satisfied results.It is of great importance to continue to study co-segmentation.Geometrical active contour models,i.e.,Level set methods,possess many advantages,such as,the unified energy functional combining the image,the intrinsic properties of evolving curve and other high-level knowledge,the ability of dealing with topological changes of the evolving curve and so on.Based on the geometrical active contour model,our thesis focused on researching the methods of image co-segmentation,and proposed two new methods,as following.(1)A PCA reconstruction error based active contour method has been proposed to realize image co-segmentation.The PCA based reconstruction technique reconstructs the original data using the linear combination of PCA orthogonal basis vectors,and employs the reconstruction error to measure the effect of the reconstruction.Our thesis introduced PCA reconstruction technique into image co-segmentation.Firstly,the basis vectors of the union of foregrounds in the pair of images and basis of background in each image are obtained by PCA decomposition.Secondly,the foreground and background of each image are reconstructed by these basis vectors respectively,and the relevant regularization items are designed based on the reconstruction error.Thirdly,the designed items are introduced into the functional of active contour model as regularization terms,and the image co-segmentation is realized by iteractively minizing the energy functional.(2)A Hellinger distance based active contour model has proposed to realize the image co-segmentation.Compared with other distance measures,Hellinger distance,measuring the similarity of two probability distributions,satisfies the rule of triangle inequality and has a simpler form.Our thesis employs Hellinger distance to measure the similarity between the foreground and background in one image,and survey the distance between the two foregrounds in the pair of the images.Based on these distances,some new energy terms are designed and incorporated into the functional of active contour model,and the image cosegmentation is finally realized by minimizing the associated energy functional.We verified the performance of our methods against three state-of-the-art cosegmentation methods on four online databases,i.e.,MSRC,images pairs,Coseg-rep and iCoseg.The experimental results show that our methods have obtained the desired results,and are superior to the representative methods particularly on complex images.