부산대학교 도서관

With the popularity of depth sensors, research on RGB-D salient object detection (SOD) is also thriving. However, given the limitations of the external environment and the sensor itself, depth information is often less credible. To meet this challenge, existing models often purify the depth information using complex convolution and pooling operations. This causes a large amount of useful information besides noise to be dropped as well, and multi-modality interaction chances between RGB and depth become less. Also, with the gradual loss of information, the hidden relationship of features between multi-level is thus ignored. To tackle the aforementioned problems, we propose a Hierarchical Transformer U-Shape Network (HierNet) that include three aspects: 1) With a simple structure, a depth calibration module provides faithful depth information with minimal loss of information, providing conditions for cross-modality cross-layer information interaction; 2) With multi-head attention, a set of global view-based transformer encoders are employed to find the potential coherence between RGB and depth modalities. With weight sharing, several transformer encoder sets comprise the hierarchical transformer embedding module to search long-range dependencies cross-level; 3) Considering the complementary features of U-shape network, we use dual-stream U-shape network as our backbone. Extensive fair experiments on four challenging datasets have demonstrated the outstanding performance of the proposed model compared to state-of-the-art models.