Based On Cross-scalre Feature Fusion And Depth Information Of Imagr Salirent Object Detection

The purpose of image salient object detection is to recognize the most important object region in the image and strip the salient region from the background.Because the salient object in the image are easily disturbed by the background region,it is very challenging to detect the image salient object.In order to improve the performance of image salient object detection,our paper studies and proposes two the image salient object detection algorithms.One is based on feature enhancement and fusion of salient obj ect detection.Firstly,convolution neural network is used to extract the features of different convolution layers of RGB image,then the high-level feature information of image is used to guide the prediction of low-level features,and proposes a high-level semantic feature to guide regression module;Secondly,based on the fact that high-level features are more sensitive to channel information and low-level features are more sensitive to spatial information,a multi-scale feature weighting algorithm is proposed by using channel attention and spatial attention to weight high-level features and low-level features respectively;Finally,the weighted features are fused and the loss training is performed to get the final saliency map.The other is based on cross-scale feature receptive and combination of salient object detection in RGBD image.Our method adds depth image to RGB image.The salient object detection based on RGBD image comprehensively considers the color,direction,brightness,texture and spatial characteristics of the RGB image,and also contains the depth information of the image.Therefore,our paper proposes a method of salient obj ect detection in RGBD image based on cross-scale feature receptive and combination,which combines the original image with the depth image.Firstly,the feature receptive module is used to expand the receptive field of features and enhance the robustness of features.Then,the cross-scale feature receptive unit is used to fuse features of different scales to improve the detail information contained in features and increase the amount of information possessed by features.Finally,the cross-scale feature combining unit is used to combine the features of different scales of depth image with those of different scales of RGB image to extract salient objects accurately.Finally,the saliency images obtained by the two algorithms are compared with the mainstream algorithms.The experimental results show that the two algorithms proposed in this paper are superior to other mainstream algorithms on the five public data sets.