Research And Application Of Salient Object Detection Based On Deep Learnin

Salient object detection is to simulate the human visual system and detect the most eye-catching areas or objects in the image,as one of the important branches in the field of computer vision,because its salient areas have important semantic information,it has been widely used in the pre-processing stage of many computer vision tasks.Early research focused on RGB images,and when faced with complex scenes,it is difficult to detect the salient regions by relying on RGB information alone.With the popularity of Depth sensing devices,the acquisition of Depth maps has become more convenient,and as the complementary information to RGB,it is helpful to understand complex scenes and promote the research of RGB-D salient object detection.In recent years,with the continuous development of convolutional neural networks,a large number of salient object detection models based on deep learning have been proposed and achieved better results,but there is still a lot of room for improvement.In this thesis,the research is carried out on the basis of convolutional neural networks,and the specific work is as follows:Aiming at how to effectively fuse RGB and Depth information,salient region localization and edge blurring,a cross modal fusion RGB-D salient object detection method with advanced semantic repair strategy is proposed.First,a dual stream network is used to extract Depth and RGB feature respectively,and then a cross modal feature fusion module is designed to fuse Depth and RGB features layer by layer and obtain six different levels of modal fusion features,using the last three layers to jointly extract richer high-level semantic information.After that,the U-shaped structure is used to fuse the feature layer by layer from the top layer to the bottom layer,and the first three low-level features are guided by using high-level semantics information before and after fusion to complete the repair of the low-level features.Through extensive experimental analysis,the method achieves excellent performance on widely used datasets,as well as greatly improves the salient object location and edge accuracy.Aiming at the current large model of salient object detection task,which leads to a poor balance between performance and efficiency,a lightweight and accurate salient object detection method is proposed.First,a dual-stream network is used to extract Depth and RGB features respectively,and for the Depth feature extraction stream,a Depth feature enhancement module is designed,which is used to process the Depth features and extract effective information before the layer-by-layer fusion with the RGB feature extraction stream to address the influence of low-quality Depth features on the fused features and obtain five different levels of modal fusion features.Then,using the semantic information has the advantage of locating salient regions and the spatial detail information can optimize salient regions,abandoning the traditional U-shaped structure of top-down fusion and innovatively divide the decoding network into spatial detail and semantic information branch,and then the first three layers of fusion features obtained from the above coding network are used to extract spatial detail features and the last three layers of fused features are used to extract semantic features,after that,a two branch fusion strategy is proposed to fuse two different levels of features in the way of feature interaction and reconstruction,.Finally,a hybrid loss function is proposed to supervise the network training to obtain accurate salient maps.The experimental results show that the method achieves excellent results while balancing accuracy and efficiency.Based on the above two methods,an image processing system for salient object detection is designed using the separation of frontend and backend technology,which separates the salient and non-salient region of the input image by embedding the salient object detection method,and intelligently achieve the rich functions of one key matting,automatic background conversion,certificate photo color change,image blur and so on.