Research And Application Of Image Co-Segmentation Based On Saliency

Co-segmentation,as a branch of the image segmentation,obtain the interested objects,which refers to partition the same or similar object from multiple related images.Co-segmentation is commonly used in object detection and object tracking,image retrieval,image classification,the 3d model reconstruction and so on.Co-segmentation is often affected by the diversity of object and complex background.This paper mainly solves two problems,which are similar scene co-segmentation in multiple images and object is similar to the partial background in the same image.The main work and the creativities are as follow:(1)Facing the situation of the similar background having negative effect on image co-segmentation,this paper proposes a method to extract saliency feature under similar scene.Firstly,the method adopts the simple linear iterative cluster to partition the image to get the content-aware superpixels of the images.Then clustering the superpixels to different classes by k-means.Basing on superpixels and cluster,we get the inter and intra contrast cue,and inter spatial cue,which are used in the process of fusion step to get the initial saliency feature.After that,using the self-organizing renew mechanism of cellular automata,we optimize the saliency based on border superpixel.Finally,images are segmented with graph cut.The experiment is performed on the Oxflord flower and icoseg dataset.The performance show better result compared with other methods in similar scene image co-segmentation.(2)Facing the problem of object similar to background and background always labeled as object mistakenly,this paper proposes an method to extract shape feature,and based on the shape feature and the proposed saliency feature build MRF model.Firstly,we use the RGB,Scale Invariant Feature Transform(SIFT)and active contour model based on closed curve to measure the similarity within the curve and outside of the curve to extract shape feature in multiple images.Then,we build an MRF energy function based on extracted shape feature and saliency feature.The complex energy function is divided into two submodules to solve the co-segmentation problem.In final,the results on icoseg and Coseg-Rep reflect the proposed model have better segmentation in images with similarity object and back-ground.(3)Based on the purpose of the co-segmentation and the existing image retrieval technology,this paper proposes an salient object based image retrieval method.And we perform the experiments on the existing flower database and the built image library.The method extracts the saliency and shape feature in the proposed co-segmentation.For the input image and images in image library,we extract the saliency and shape feature,and measure the similarity between the input image and library images.Then output the retrieved similar images.