Research On Salient Object Detection And Lightweight Method Based On Depth Information Fusion

Salient object detection aims to identify and locate the most attractive regions in an image by simulating human visual attention mechanism,which has been widely used in various vision tasks such as image recognition,object tracking,image quality evaluation and camera focusing.In recent years,RGB salient object detection has developed very well.However,the performance of detection is still poor under complex background conditions.With the development of RGBD sensor technology,it can capture the RGB information in the scene and obtain the corresponding depth information at the same time.Depth cues provides more spatial clues for salient object detection,which enhances the model's ability to distinguish the difference between the salient object and the background region.Most of the existing salient object detection models are based on deep convolutional networks which need a lot of data for training.However,the small scale of RGBD datasets are not conducive to the learning of the deep model,and the large amounts of parameters of the model could not meet the needs of practical applications.Therefore,it is of great significance to improve the learning ability of the salient detection model based on convolution network for small-scale RGBD data and the lightweight transformation of the model.The main work of this article is as follows:(1)Research on depth fusion salient object detection.We use the transfer learning method to transfer the knowledge learned from large-scale RGB data to RGBD salient object detection,we found that leaning ability of model is enhanced;Through the research of deep fusion network,it is found that the method of hierarchical feature fusion is more conducive to the complementary expression of RGB and depth information,and improves the accuracy of model.Therefore,combined with the method of transfer learning,we propose a salient object detection model based on hierarchical feature fusion.(2)Research on the lightweight of RGBD salient object detection model.We use the depth separable convolution to replace the standard convolution of the RGBD salient object detection model,and find that the parameters of the model are greatly reduced;In the research of residual connection,we found that the difference feature of convolution input and the output is used to assist the backbone network can improves the detail expression ability of the model;Therefore,on the basis of deep separable convolution,we design a new residual connection and propose a lightweight model of RGBD salient object detection with parallel multi-scale residual structure.