The Research And Application Of Co-Saliency Detection Based On SPCA

Co-saliency detection is a kind of technology aims to detect the similar objects in multiple images. Co-saliency detection has gained increasing wide applications such as co-segmentation, information retrieval, automatic annotation and foreground extraction, and gets more and more attention from various researchers. Although several co-saliency detection algorithms were proposed in recent years, how to achieve fast and robust detection still needs further study in complex environment. Focusing on the deficiencies of the existing co-saliency technologies, this paper makes some research on fast co-saliency detection and its application based on sparse Principal Component Analysis.In this thesis, a fast co-saliency object detection algorithm based on sPCA is designed by a step by step process. First of all, we achieve the fast preprocessing step based on low level feature extraction. By considering the importance of effective features, feature selection in done to improve the accuracy of the K-Means co-saliency detection algor ithm. Finally, we propose an accurate co-segmentation algorithm based on co-saliency detection. The main works as follows:(1) In order to improve the efficiency, an algorithm based on block-based Truncated Power method as a preprocessing step is proposed. The experimental results prove that this improvement eliminate the massive redundant data and retain the effective image feature to shorten the running time of co-saliency detection.(2) An improved sPCA method with feature selection through the effective distance between the blocks is presented to solve the loadings inconsistency in various principal components of sPCA. The experimental results demonstrate that our method can improve the efficiency of K-Means co-saliency detection while ensuring the detection accuracy.(3) We research on the segmentation technology based on co-saliency detection algorithm. An automatic and accurate salient object segmentation method is proposed by comprehensive considering the advantages of co-saliency detection, graph cut segmentation with star prior, one-step GrabCut and active contours methods. The experimental results show that it can obtain the accurate and robust segmentation results.