Research On Detection And Retrieval Methods For Salient Objects

Salient object detection is a research hotspot in the field of machine vision.In recent years,many computational models have been developed for salient object detection.In this paper,we formulate a new saliency detection model,in which saliency map is computed as a combination of the colour,its distribution-based saliency and orientation saliency.Then we apply saliency detection to content-based image retrieval.The main content of this paper is as follows:(1)We analyze several classical saliency detection methods and their advantages and disadvantages,and an unsupervised saliency detection method based on reciprocal filter is proposed to solve some problems.Due to the salient regions detected by SF method lack of integrity,the proposed method fuses the local orientation information(LOI)in Itti model with colour uniqueness and its distribution used in SF method to make up the above deficiency.Meanwhile,we define a reciprocal function to substitute the Gabor function used in LOI to detect relatively homogeneous and uniform regions at the boundary of salient object.Experiments show that our method outperforms state-of-the-art methods on four benchmark datasets.(2)A salient object-based image retrieval method(SOBIR)is presented,which extracts features of the salient objects contained in target and candidate images respectively to carry out contentbased image retrieval(CBIR).Our method utilizes colour histogram in the L*a*b* colour space as colour represetation of the salient object,and the shape representation is extracted from the binary image using an improved polar Fourier.The similarity between the two salient objects contained in the target and condidate image is defined based on the linear combination of colour and shape representations.Experimental results show that,compared to the latest two CBIR methods,the proposed SOBIR method exhibits an excellent performance in precision,recall,flexibility and efficiency.