•  
  •  
 

Journal of System Simulation

Abstract

Abstract: In order to solve the problems that the boundary of the saliency object detection area is vague, and the detection area is incomplete or inaccurate, an RGB-D saliency object detection method based on cross-refinement and circular attention is proposed. A cross-refinement module is designed at the stage of extracting features using encoders, which is used to supplement feature information of each other and improve the feature quality before fusion. It also suppresses the negative impact of poor-quality depth maps and addresses the issue that the edges of the saliency object are blurred. For the features after fusion, the circular module is proposed, which combines the attention mechanism with convolutional long short-term memory (LSTM) network unit to simulate the internal generation mechanism of the brain and help infer the current decision by retrieving past memories, so as to obtain semantic scenes that require long-term memory. The module can comprehensively learn the internal semantic relationships of fusion features to generate a more complete and accurate saliency map for the detection area. Experiments conducted on six public datasets show that the proposed method can obtain a saliency map with clear edges and high accuracy.

First Page

1931

Last Page

1947

CLC

TP391.9; TP391.4

Recommended Citation

Dong Qingqing, Wu Hao, Qian Wenhua, et al. RGB-D Saliency Object Detection Based on Crossrefinement and Circular Attention[J]. Journal of System Simulation, 2023, 35(9): 1931-1947.

Corresponding Author

Wu Hao

DOI

10.16182/j.issn1004731x.joss.22-1372

Download

Share

COinS