PA2Net: Pyramid Attention Aggregation Network for Saliency Detection

Jizhe Yu, Yu Liu, Xiaoshuai Wu, Kaiping Xu, Jiangquan Li

Abstract: With the development of deep convolutional neural networks (CNNs), salient object detection has become increasingly mature. Existing methods primarily enhance model performance by deepening and widening U-shaped networks. However, these methods typically consider the top-level features of bottom-up processes as global context, which is suboptimal and limited. This limitation arises because existing CNNs do not consider long-range dependencies, focusing more on local convolution operations. Additionally, the implicit interaction between complementary features in the top-down process fails to effectively learn the structural properties and semantic information, and it inevitably introduces significant noise. To address these issues, this paper adopts the idea of attention mechanisms to restore the object integrity in a more explicit manner while avoiding the introduction of irrelevant factors. First, we design a Global Context Enhancement (GCE) module to further enhance object localization and global context modeling by learning the correlations among deep features. Subsequently, an attention-based Pyramid Aggregation Decoder progressively merges multi-level adjacent features by learning their pairwise correspondences, gradually learning local details and global representations. Furthermore, a Noise Filter Block (NF) is integrated into the PAD to suppress background noise interference. Experimental results on five benchmark datasets show that our model performs favorably compared to 15 state-of-the-art methods.

External IDs:dblp:conf/mmm/YuLWXL25