Saliency detection combined with selective light field refocusing of camera array

For the multiple salient targets scene, as well as a scene in which some areas of the salient target do not contrast significantly with the background area, the saliency maps obtained by existing algorithms are not fine enough or even lose some saliency regions. In this paper, a new saliency detection method combined with selective light field refocusing of camera array is proposed. In this method, the light field dataset is selected and multi-viewpoint images of the same scene are used. First, we perform refocusing rendering combined with super-resolution on the central viewpoint image. Then, on the basis of the graph-based saliency detection method, we propose a propagation model combining global and local smoothness constraints to prevent false label propagation. Finally, the obtained coarse saliency map is refined through the object map to output the final saliency map. In addition, for the scene that contains multiple salient targets, by refocusing a certain depth layer in the scene, and producing varying degrees of blurring to other depth layers, the salient targets on the depth layer can be detected accurately and in detail. To a certain extent, the optional saliency detection is realized. Experiments on the 4D light field dataset show that the average Mean Absolute Error (MAE) between the ground truth and the saliency map obtained by the method proposed in this paper is 0.2128, which is lower than those obtained by the existing methods. The detection result contains more detailed information about the salient target, which improves the above-mentioned shortcomings of the existing salient detection methods.