Image Saliency Detection Based On Bayes Integration

Computer vision can simulate human visual mechanism,whose function is similar to the function of human brain.It can automatically screen,understand and analyse the main information in natural situations.As the core of computer vision research,saliency detection has widely attracted the attention of researchers.Saliency detection has been applied for many areas,including image segmentation and target recognition.According to detection principle,saliency detection is screening out main information at images in a short time when the visual range is wide,which greatly improves the rate of image processing.Because of the varieties of images,the detection of the image salient region is gradually improving.Aiming at the problem of inaccurate saliency detection and unclear edge in current image saliency detection algorithms,an image saliency detection algorithm based on improved Bayesian integration is proposed.At first the image is segmented into superpixels and the boundary is removed,then the dark channel prior and compactness prior features are extracted.Salient objects are detected and primary saliency map is obtained.Also,the multi-kernel learning is used for the primary saliency map to train and optimize,and the strong saliency map is obtained by smoothing the salient objects edge with the guide filter.Finally,the integration of primary saliency map and strong saliency map is completed based on improved Bayesian,and the final saliency map is obtained.The main works of this thesis are as follows:At first the image is preprocessed,the image is segemented into superpixels by using SLIC.According to the principle of photography,there is usually no salient region at the boundary of the image,so removing image boundary,which can make the algorithm calculation more concise and reduce the time.In the process of constructing the primary saliency map,it is usually calculated directly by using features such as color,although the calculation speed is fast,the result is relatively poor.In this thesis,the method of obtaining the primary saliency map is proposed in which color,brightness features and image space compactness are combined,the dark channel prior and compactness prior features are extracted.In order to further solve the problem of large number of images,different image sizesand unclear classification of image objects in datasets,multi-kernel learning method is used to optimize the primary saliency map.The training and testing are carried out on an image,then the guided filtering method is used to further optimize and smooth the edges of salient regions to obtain the strong saliency map.Because the size of objects in the image is difference,salient objects may be at different scale,but traditional single-scale segmentation methods can not flexibly control the scale and are extremely sensitive to the number of superpixels.Therefore,multi-scale superpixels segmentation and features extraction are used to obtain the primary saliency map which can highlight the salient object uniformly.In order to make full use of the advantages of primary saliency map and strong saliency map,the primary saliency map and strong saliency map are taken as prior probability and likelihood probability respectively,two posterior probabilities are calculated.Different weights are used to integrate the two maps to achieve the Bayesian integrated and the final saliency map is obtained.The proposed algorithm is compared and evaluated on 4 public datasets with the results of 8 popular algorithms,and qualitative evaluation and quantitative evaluation indicators(Precision,Recall and F-measure value)are analyzed.Experimental results show that the proposed algorithm can better suppress background and improve the accuracy of the detection results.