Research On Co-Saliency Detection Method Of Single Image

The visual attention mechanism is a very important and practical biological mechanism in the human visual perception system.It can help humans to quickly search and locate areas of interest in complex natural environments.The saliency detection is based on the study of this biological mechanism and the simulation of the mechanism,so that the computer can quickly find the region of interest in a given image,that is,the saliency region.With the development of computer networks and the upgrading of camera equipment,not only has the number of data images exploded,but the information contained has also become more abundant.Under such circumstances,co-saliency detection has also gradually developed.Co-saliency detection on a single image as one of the new research directions.The areas marked as foreground are no longer all the saliency areas in the image,but the salient objects with high similarity in the image.Compared with saliency detection,there has an implicit condition for images,that is,at least two salient objects are contained in the image and appear as co-saliency.Co-saliency detection has a high value in further reducing image information redundancy,similar objects statistically,and animation synthesis.Actually,the co-saliency refers to the fact that two salient objects are identical or similar in a certain feature,so as long as the feature can be determined,the co-saliency region can be obtained by clustering.However,the scenes and contents displayed by different images are not the same,to obtain the best results,the features used for clustering should be the most suitable.In order to solve this problem,this thesis regards feature selection as an optimization problem and proposes a single image co-saliency detection algorithm based on particle swarm optimization optimal clustering.In the first stage,according to the difference in the proportion of the co-salient part and the non co-salient part in the foreground,the Bayesian formula is used to calculate the co-salient probability of each superpixel to generate a rough co-saliency map,as the initial saliency map in second stage;then in the second stage,extract 7 kinds of 20-dimensional feature vectors for superpixels at first,and then the PSO algorithm is used to select the appropriate clustering feature for each image in the process of multiple iterations,and finally the feature is used to generate the final co-saliency map.In the second work of this paper,based on the idea of directly comparing the features of each object in the image to judge the co-saliency,a simple and efficient single image co-saliency detection algorithm is proposed.First,generate the object boxes by using the existing object detection algorithm.And then,in order to ensure the accuracy of co-saliency detection,extract the unique area of each object by combining the initial saliency map and the overlapping relationship.Then extract the feature of each object and calculate the similarity,mark the cosaliency objects initially.Finally,the support vector machine is used to classify the superpixels in the image,and the co-saliency value is assigned to the entire image to obtain the final cosaliency map.In summary,the two single image col-saliency detection algorithms proposed in this paper are tested on the WICOS dataset,and the detection results are compared with other methods.Experimental results show that these two methods have certain advantages in PR curve,maximum F value and mean absolute error.