Automatic Natural Image Segmentation Based On Saliency And Interactive Segmentation Algorithm

In nowadays world while computer network, multimedia technology and imagingdevices developed rapidly, mass image data is going through explosive growth. So how touse computers to process image data more intelligently and serve people better becomesextensive demand and hot research issue. But during detecting salient regions, theChallenges are mainly caused by natural images containing complex scene. At the sametime, unsupervised image segmentation techniques are also flawed while dealing with thesame problem. Because no semantic information is given, the regions of interest are hardto be detected. On the other hand, interactive segmentation algorithm may bringuncertainty and low effectiveness because the step of manual intervention.According to the obstacles just mentioned above, this thesis proposed an automaticnatural image segmentation algorithm based on saliency and interactive segmentationalgorithm to segment ROIs from images. The Proposed method has advantages overtraditional methods because: saliency offers priori knowledge which can lead to moreprecise and intact ROIs segmentation; On the other hand, leaving out manual interventioncan help avoid the indeterminacy that subjective information may brings. The majorfactors that affect the proposed method’s result are the saliency map’s precision andintegrity, and the veracity of the priori information given to the segmentation step. So, theprimary research points of the thesis are:(1)Summarized and concluded existing saliency, segmentation methods both homeand abroad. Proposed a saliency evaluation method based on both global contrast andregion contrast to obtain saliency map that more accord with human visual attentionmechanism and more suitable for segmentation step.(2)Via processing the saliency map, the thesis obtained image labels which arehelpful for the segmentation step. Instead of traditional state-of-art image segmentationalgorithm, the thesis applied the obtained saliency labels to replace the interactive labelingstep of growcut algorithm, to realize both precise and intact image segmentation.Using the method proposed in the thesis, saliency map and segmentation result areobtained and tested in the standard database, MSRA Salient Object Database, and bothoutperformed state-of-art methods in subjective visual effects and objective indicators. It is demonstrated that our methods improve the ROIs segmentation results.