Research On Image Saliency Detection Algorithm

In today's society,information science technology has been developed rapidly.People accept all kinds of information proactively or passively in everyday life.However,today's society moves forward at a fast pace,this requires people to extract useful information quickly in a large amount of information.Among the various kinds of information people accept every day,the proportion of image information is very large.So in the field of computer vision the technology of image information processing,especially the extraction of salient information in a picture is of significant importance in building a convenient information age.This paper systematically introduces the basic theory of image saliency detection and studies several classic saliency detection algorithms.Aiming at solving the problem that the target edge is not clear and the background suppression is not enough in the salient map produced by the traditional algorithm.A saliency detection method based on histogram contrast and guided filtering and a method of saliency optimization based on background prior are proposed.The effectiveness of the algorithm proposed in this paper is verified by experiments.The paper has the following main work and contribution:1.Several classical algorithms in the field of image saliency detection are deeply studied,meanwhile the principle of the algorithm is given,and their advantages and disadvantages are analyzed in detail.2.A saliency detection algorithm based on histogram contrast and guided filtering is proposed.First,we use the method of histogram-based contrast(HC)to detect salient objects,thus get a preliminary saliency map;Secondly,through the edge extraction and background contour elimination of input original image,we can get the outline map that can roughly frame significant objects,then guided filter is conducted to get second initial saliency map;Finally,based on regional energy,two initial saliency maps are fused according to specific rules to generate the final saliency map.The saliency map obtained by this algorithm not only keeps the good ability of HC map extracting internal information of salient objects but also possess good retention of the edge of a significant target.The objective evaluation shows that the performance indicators of the algorithm proposed is more prominent than those of commonly used saliency detection methods.3.A saliency detection algorithm based on background prior is proposed.Image edge extraction is the basic step of image understanding and analysis.This method uses Convolutional Neural Network(CNN)to obtain the edge features of the image.Owing to the characteristics of convolution and pooling calculation of convolution neural networks,the translation of images will not affect the final feature vectors.Therefore,the probability of over fitting of image features extracted by CNN is very low.What's more,the process of extracting features by CNN is more scientific.Meanwhile,image features extracted by CNN are more accurate.Then,the edge map is applied to the saliency detection algorithm based on robust background detection,thus the first initial saliency map is obtained.Next,the Simple Linear Iterative Clustering(SLIC)super pixel segmentation algorithm is applied to the input image,and the saliency of the image is calculated with the specific super pixel as the background prior.Finally,the second initial saliency map is obtained by using different fusion rules in different regions.Experimental result proves that the algorithm can not only detect the salient target accurately,but also suppress the image background better.