| Salient object detection is designed to detect and segment the most striking objects in an image.It is a basic preprocessing task in computer vision and can be applied to various visual tasks such as target tracking,object recognition,image classification,and so on.RGB-D salient object detection can alleviate the problem that RGB images are susceptible to environmental lighting and other factors by introducing extra depth images(Depth)to supplement RGB images with spatial information.Although RGB-D salient object detection has made great progress with the rapid development of in-depth learning,it still faces the problems of inefficient depth feature extraction,poor feature expression ability and adaptability of network models.Therefore,this paper studies efficient depth feature extraction and establishes a universal multimodal salient object detection model.The main research work of this paper is as follows:(1)This paper first proposes a guided residual network model based on efficient depth feature learning.First,a simple depth feature extraction branch is designed to replace the backbone network of the existing dual-stream detection network.The depth feature is weight corrected by depth correction module to reduce the negative impact of low-quality depth map on the network model.Then,the depth feature is extracted by convolution and residuals overlay.Secondly,the multi-scale receptive field module is embedded in the deepest part of the network to expand the receptive field.Finally,image features and depth features are progressively fused through the residual thinning module.(2)This paper further presents a split-guide network model based on efficient depth feature learning.First,this paper uses the overlay of ordinary convolution and pooling layers to extract depth features efficiently,which is more efficient and concise than the depth backbone network in the existing RGB-D dual-stream structure.Then,from shallow to deep,the side output fusion features are obtained by concatenating depth features and image features layer by layer.Finally,a split-guided convolution module is proposed,which can effectively enhance the characterization ability of convolution features and expand the perception field of fusion features.The network structure is robust and can be used for other salient object detection tasks such as RGB,RGB-T,etc.(3)The method proposed in this paper is compared with existing advanced methods in the RGB,RGB-D,RGB-T salient object detection field on public datasets.A large number of experimental data show that our method obtains the best performance on the four universal evaluation indexes at that time.In addition,the network model proposed in this paper has more advantages in efficiency,parameters and model size.Finally,the proposed model is applied to RGB-T significant object detection,RGB-D portrait segmentation,RGB-D disguised object detection,and good performance is obtained. |
