Abstract:
In order to improve the detection accuracy of computer vision systems for salient objects in visual scenes, an adaptive sampling visual saliency detection method incorporating background prior information is proposed. We use superpixel segmentation to extract image edges to obtain scene prior information. Then, the frequency domain Gaussian difference feature spectrue spectrum is mapped to the spatial domain to generate a gray density scatter plot. Next, the sampling windows adaptively move to the foreground region according to the scatter plot to achieve adaptive sampling. The process simulates human eye tracking to detect salient objects. Finally, We use the principal component analysis method to fuse background prior information and adaptive sampling features to synthetically extract foreground information. The saliency map with higher resolution is obtained. The experimental results show that the adaptive sampling method is efficient in comparison with other sampling mechanisms, and the proposed algorithm is compared with 13 other algorithms on ECSSD, HUK-IS, MSRA-5K, and SOD public datasets. On each dataset, MAE decreased by an average of 0.01−0.04, F-measure increased by an average of 0.01−0.04, and IoU increased by an average of 0.02−0.08, which verify that the proposed algorithm has state-of-the-art in the accuracy of salient object detection.