Depth Quality Enhancement and Feature Shrinking Guided RGB-D Salient Object Detection

Nowadays, depth cameras are equipped on many mobile devices, meanwhile RGB-D based salient object detection methods have made significant progress based on deep neural networks. However, poor quality depth images are not avoidable, which can certainly degrade the performance of current salient object detection method. Therefore, it is worthwhile to mitigate the impact of low-quality depth images while maintaining a lightweight model. This paper proposes a depth quality enhancement(DQE) and feature shrinking(FS) guided salient object detection network, named DEFSNet, which consists of RGB and depth streams based on MobileNet-V3 and depthwise separable convolution. The model also includes a Dice coefficient-based depth quality judgment(DQJ) mechanism to assign weights to each depth feature and a depth feature enhancement (DFE) structure to improve low-quality depth images. Finally, a feature shrinking structure is designed to fuse multi-scale features and output the saliency prediction map. Experimental results show that, compared to state-of-the-art salient object detection models, the proposed model achieves higher detection accuracy with a smaller number of parameters.