Research On Segmentation-based Image Saliency Detection

Image saliency detection is a basic problem in the field of computer vision and image analysis.Related algorithms have been widely used in object recognition and tracking,image retrieval,image segmentation,data compression,and so on.In this paper,several key problems in the existing algorithms are pointed out,and then a segmentation-based strategy for saliency detection is developed and exploited.In consequence,more reasonable ideas for conducting salient region detection are proposed and several new algorithms are developed.Existing saliency detection algorithms can be divided into spatial-domain-based algorithms and frequency-domain-based algorithms.It is observed that spatial-domainbased algorithms are suitable for processing images with color saliency,while frequencydomain-based algorithms are suitable for processing images with non-color saliency.By using an image classification technique,we proposed an algorithm that could be used to process the two above-mentioned types of images in the first part of this paper.Firstly,input image is classified according to their saliency types.Then,its saliency map is computed.Finally,the saliency map is further enhanced by using a saliency cut method.Experimental results show that our new algorithm is comparable to the existing saliency detection algorithms.It is observed that the existing saliency detection algorithms usually use a center-bias assumption.To be specific,it is assumed that salient objects always appear in the center of the input image.However,this is not always true,especially for those images acquired by unmanned monitoring system or device(e.g.,surveillance camera),in which the salient object could appear in any location within the image.Consequently,the resulted saliency detection performance could be greatly degraded.For example,the saliency detection algorithm based on global contrast incorporates the center-bias setting in the process of computing saliency map.To remove this unreasonable assumption,a location-aware strategy is proposed to identify the location of salient regions preliminarily in the second part of this paper.As a result,a location-aware algorithm is proposed based on this strategy.Extensive simulation results show that the proposed algorithm clearly outperforms the existing stateof-the-art algorithms.Although the detection accuracy of deep-learning-based saliency detection algorithms is better than traditional algorithms,it could be further improved.In the third part of this paper,a saliency detection method based on fully convolutional networks is proposed.Firstly,we compute an initial saliency map by using deeper network and atrous convolution.Then,fully connected conditional random field is adopted to optimize the boundary of saliency map.Finally,a saliency cut technique is used to further enhance the saliency map.Experimental results demonstrate that our proposed method is superior to the traditional saliency detection algorithms and the existing deep-learning-based algorithms.The above-mentioned three new algorithms for image saliency detection have different characteristics in terms of technical route and practical effect.In this paper,a comparison between these methods has been conducted and the application scenarios of each algorithm have been clarified.