Research On Salient Object Detection Method Based On Multi-Feature Fusion

Facing the huge amount of data in the era of big data,it is necessary to use limited computing resources to obtain the maximum information.Salient object detection aims to imitate the human visual attention mechanism to detect and segment the salient object in the scene that attract the most visual attention.In many object-level application fields,such as image cropping,video salient object detection,and visual tracking,salient object detection plays an important role as a precursor work and is a hot research direction in image processing work.In recent years,benefitting from the rapid development of deep learning,many salient object detection models based on convolutional neural networks are proposed,making breakthroughs in this field.With the gradual saturation of network models in terms of performance,the focus of salient object detection has gradually shifted from accurately locating salient regions to refining the structure of the object,including the integrity of the object as well as sharpening the boundary of object.However,the main problem of the existing models is that with the deepening of the network,the shallow details of the network are lost,which makes it difficult for the models to effectively learn the boundary region information and context details of the salient objects.In addition,for all kinds of features extracted from each layer of network model,there is a lack of appropriate fusion methods.In view of the current difficulties,this paper focuses on solving the core problems of multifeature extraction and fusion,and at the same time considers the modeling of object boundary,aiming to build a complete,operational and verifiable salient object detection model that can adapt to complex scenes.The main research contents are as follows:(1)Propose a salient object detection method based on the fusion of content and edge features.For the predicted object structure incompleteness problem,a content-aware feature extraction module is designed,which uses a self-attention mechanism and a channel-attention mechanism to capture content information from both spatial and channel dimensions.For the problem of incomplete boundary of the predicted object,an edge-aware feature extraction module is introduced to learn boundary features and predict the complete boundaries of the object.In addition,a learning-based feature fusion module is proposed to take the features in the above two modules and fuse the features and complement the information at each stage in a learning manner.(2)Propose an RGB-D salient object detection model based on axial attention and multimodal feature fusion.The depth image,as a complement to the RGB image information,can effectively reduce the impact on the input information due to noise such as illumination.In the model,the RGB image data is feature-enhanced using the axial attention mechanism and learns boundary features through the embedded edge feature guide module;the depth image data is feature-enhanced using the multiscale feature extraction module;and the RGB image and depth image features are integrated and enhanced by the adaptive multi-modal feature fusion module and achieve complementary multimodal feature information.Based on the common benchmark RBG dataset and RGB-D dataset,the proposed methods are compared and analyzed with the state-of-the-art salient object detection methods according to the commonly used evaluation indexes,and both methods in this paper have improvements to some extent.