Research On Image Retrieval Method Based On Salient Object Detection

Image saliency object detection based on the mechanism of human visual attention detects the regions which attract the person's visual attention in the image.It has been widely applied in image retrieval,object detection,image compression and semantic annotation.Although image saliency object detection has achieved great success for more than 10 years,it is still a challenging research,and requires more effective methods to deal with more complex images of the background and foreground.In most of the current saliency detection methods,the detection result contains most of the background information or loses some information of saliency object.To solve this problem,this paper proposes two models to extract a more complete salient region,one is a saliency detection model based on neighbourhood,the other is saliency region detection model based on region merging.These models of image saliency object detection are used for image retrieval by measuring the feature similarity between salient regions that contains important content.1.We propose a saliency detection algorithm based on neighborhood optimization mechanism.Firstly,we divide the input image into a number of super pixel blocks to improve the computational efficiency of the algorithm.Secondly,we apply the contrast and distribution characteristics of different regions to construct the contrast map and distribution map,and merge two maps through a new merge function.Finally,we define the neighbor relationship between super pixels,and establish an update optimization mechanism.In the iterative update process,the saliency state of the super pixels is determined by its current neighborhood and itself,which can reduce the misallocation of regions.It is tested on two standard image databases and compared with current classical saliency detection algorithms.The results show that the proposed algorithm has better detection effect and higher accuracy.2.We propose a saliency detection algorithm based on region merging.Firstly,according to the theory of most of the boundary regions belong to the background region,we construct an initial saliency map by clustering with boundary pixels.Secondly,we compute the integrated values of region and merge the adjacent regions by ranking them in the first stage of the region merger.Finally,in the second stage of theregion merger,we construct a final saliency map by merging the nonadjacent or adjacent regions,which is based on the area of region,and distant from image center and utilizes the length of boundary.It is tested on three standard image databases and compared with current classical saliency detection algorithms.The results show that the proposed algorithm has better detection effect.3.The saliency detection algorithm is used for image retrieval by extracting the salient regions and measuring the similarity between salient regions based on color moments and LBP textures.From the final simulations,we can demonstrate that the image retrieval based on the salient region is more accurate.In this paper,we construct two saliency detection models with a higher accuracy to extract the object which can attract the visual attention in the image.Ignoring the unimportant information of the image background and extracting the salient regions that contain the complete information,it can improve the accuracy of the image retrieval by the simulation test.