A salient object detection algorithm based on RGB-D images

To solve the problem lack of RGB-D dataset for training, a salient object detection algorithm by cross dataset training only using RGB dataset is proposed. First, a simple convolutional neural network is designed to prediction foreground and background trained on RGB dataset MSRA10k. Then, the SLIC superpixel segmentation method is used to fuse the depth image information and cluster pixels, which can segment the edge of salient object more accurately. Finally, based on the global distribution characteristics of salient objects, superpixels are labeled using kernel probability density estimation. In order to verify the effectiveness, the proposed algorithm is compared with three newer algorithms, which has achieved better detection results in terms of PR curve, AUC and F-Measure. Experimental results show that the proposed method can improve the salient object detection of RGB-D image in the absence of RGB-D images for training.