Salient Object Detection based on Adaptive deep Differential Pyramid

In the initial stage of salient object detection, the main structure of the saliency model is based calculation of Center-Surround Contrast, which is introduced into a few CNN-based saliency detection model. However, the receptive field of feature layers could not be dynamically adjusted in these methods. Therefore, an Adaptive depth Difference Pyramid (ADP) module is constructed, in which receptive fields could be dynamically adjusted to calculate the difference between features of different levels, sensing the local characteristics of saliency objects. Subsequently, an ADP-based method for salient object detection is proposed. Experiments was conducted on 4 public datasets, and the mean absolute error (MAE) and maximum F-measure was selected for the evaluation metric, which achieved MAE of 87.1%, 86.7%, 88.9% and 82.3% respectively.