Research On Image Saliency Detection Algorithm Based On Multi-priori Fusion Strategy

The image saliency detection is one of the hot topics in the field of machine vision,which imitates human visual attention mechanism by the computer,to extract the significant areas in the image.The image saliency detection is deeply studied in this thesis,focusing on the extracting the salient object accurately,ameliorating the ambiguous phenomena around the salient object and enhancing the applicability of the algorithm,the saliency detection algorithm based on multi-priori fusion strategy and its optimization algorithm are proposed.The main contents are as follows:(1)The fusion propagation and saliency detection of multi-priori labelsThe more accurate characterization of image organization and content can be achieved,based on the available deep-level effective visual priori features.Therefore,the object priori label and background priori label can be extracted separately by fusing the center's priori,objectness priori,boundary priori and edge probability priori.Two types of labels which incorporate multiple priori features are integrated in a global computable multi-label propagation framework in this thesis.The salient information carried by the labels is spread accurately through the updated principle of cell machine.Experiments show that this algorithm can capture salient object successfully.(2)The local analysis of salient object contour priori labelsThe ambiguous phenomena occurs because the background pixels around the salient objects are usually given higher saliency easily.And so the contour of the salient object is not clear.In order to ameliorate this phenomena,a new salient object contour priori label is defined,and a novel algorithm based on analyzing local of salient object contour priori label is proposed in this thesis.In the local scope,the boundary line is constructed according to the contour prior label,the relative position relationship between the center of contour labels and local superpixel is analyzed and discussed.Superpixels in the scope of salient and background are categorized and marked carefully again,in order to determine whether or not the superpixel belongs to the salient object area.Experiments have proved that the proposed algorithm can ameliorate the ambiguous phenomena around salient object effectively.(3)The guidance optimization of multi-priori saliency resultsFor some image data with complicated background and various structures,the priori information descriptions of the image content and organization sometimes deviate from the human visual perception.The multi-priori fusion algorithm mainly relies on the effective priori to detect salient object.It is unable to obtain satisfactory saliency detection results in some cases.In order to establish a saliency model with strong applicability and complete extraction,under the guidance of the saliency detection results,a classifier with better classification performance is constructed to separate the salient object from the background area,combined with the theory of multi-core boosting learning algorithm in this thesis.Due to the introduction of the global image visual clue,experiments have proved that the deficiency of priori describing image feature in some cases is made up,the overall performance of the proposed saliency detection algorithm is further improved.