Research On RGB-D Image Salient Object Detection Algorithm

In the real world,humans quickly search and perceive attractive areas through the visual system to extract key information.In the era of big data,electronic devices have brought massive amounts of image data,and computer technology has been used to simulate the human eye's visual attention mechanism,which is the research content of significant object detection.In computer vision,salient object detection is a key pre-processing step.By focusing on salient objects to reduce the workload of calculation,this is widely used in image retrieval,image segmentation,object detection and recognition and other technical fields.In the area of salient object detection,there are currently two research directions: salient object detection of a single image and salient object detection of multiple images.In recent years,the saliency detection methods of RGB images have developed rapidly.These methods are for 2D image processing,but saliency detection also has practical application scenarios in 3D vision.With the emergence of RGB-D images,depth information is gradually being mined.People use color information and depth information to automatically locate objects of interest to the human body,which reduces the complexity of visual analysis.Research on significance detection of RGB-D images is also favored by more people.According to the above,salient object detection can be further divided into RGB single image salient object detection,RGB-D single image salient object detection,RGB collaborative significant object detection,and RGB-D collaborative salient object detection.The research direction of this article is RGB-D Single image salient object detection.For the saliency detection task of RGB-D images,how to obtain discriminative depth features is the first key issue,and the second issue is how to combine low-level features with depth features more effectively.On the other hand,with the in-depth study of Machine Learning,machine learning methods can divide data,train excellent models,and use model prediction to separate foreground and background.However,these good classifier methods are mainly used in the field of RGB salient object detection,but have not been well developed in the field of RGB-D salient object detection.Based on the above problems,the work of this paper has two aspects.This paper proposes a background and foreground fusion RGB-D saliency detection model,which takes depth information into consideration to extract salient regions from both the background and foreground.Firstly,the background measurement mechanism of the image boundary information was used to remove the foreground noise and selected the background seeds from the boundary superpixels calculating the background-based saliency map.Secondly,the input image was constructed into a map and the depth information was added into the graphics structure.The clues,including color,depth,and position,were used to obtain the foreground seeds,calculating the foreground-based saliency map.Finally,the background and foreground image were fused to obtain the initial saliency map.In addition,cell optimization and iterative propagation were performed to get the final saliency image.Comparative experiments were performed on the RGB-D data sets LFSD,NJU400,and NJU2000.Results showed that the propose method is effective and can improve the accuracy.Since traditional methods mainly use artificially designed low-level features to perform saliency detection,In this study,an RGB-D saliency detection method through multiple instance learning was proposed by studying machine learning.Firstly,extract image features,perform multi-scale superpixel segmentation on the image,and extract the corresponding RGB features and depth features.Secondly,obtain weak saliency maps based on the depth features and dark channel features,and separate the weak saliency maps into positive and negative samples.The training set is classified by the key example support vector machine(KI-SVM)to obtain the initial saliency map.Finally,the multi-scale weak saliency map is fused with the initial saliency map to obtain the final result.Evaluation experiments are performed on the RGB-D data set NLPR1000.The experimental results show that the model in this paper has obtained good results.