Salient object detection based on adaptive difference of Gaussian feature pyramid network
Xinrui Zhang, Jin Chu Wu, Lei Zhu · 2024
The classical center-neighborhood contrast theory, based on a biologically sound architecture, is widely used in the field of salient object detection under traditional methods, but has rarely been introduced into deep convolutional neural networks (CNNs). To take advantage of this biological property, this paper proposes a network that fuses center-neighborhood contrast with the help of a difference of Gaussian (DoG) feature pyramid. The existing difference of Gaussian pyramids utilize predetermined Gaussian kernel sizes that remain fixed throughout the training phase and therefore may not accurately measure the true pop-out attribute of the salient object. Based on such a premise, this paper proposes a Gaussian difference feature pyramid network with adaptive kernel properties, which takes the Gaussian kernel size as a learnable weight, and with the help of the gumbel_softmax's derivability, achieves the dynamic adjustment of the standard deviation of initial Gaussian kernel. Comparison experiments are conducted between the proposed model and eight popular saliency detection algorithms on five common datasets, taking the mean absolute error (MAE) and the maximum F-measure as evaluation matric. The experimental results show that the proposed model has achieved better detection performance for salient objects than the competing approaches.