Research On Visual Saliency Detection And Application Based On Deep Learning

With the rapid development of Internet technology and the advent of the era of 5G,all kinds of multimedia data such as pictures and videos have become an important means of information expression,and the popularization of interactive media is also accelerating,so it is of great significance to extract salient areas in images efficiently.Visual saliency detection refers to the accurate prediction of the significant objects in the scene through feature analysis and the acquisition of the boundary of the salient objects.The existing methods have the problems of low accuracy and slow speed,which cannot be well applied to the practical application scenarios.Therefore,this thesis adopts deep learning technology to conduct research on visual saliency detection methods,including object-level edge detection,salient area detection and its related applications.The main work and results are as follows:(1)Aiming at the problem of rough edges in salient regions,an object-level edge detection method based on multi-scale residual network is proposed.Firstly,on the basis of the deep residual network,a hybrid dilated residual module is designed to increase the receptive field of the network.Then,a multi-scale feature enhancement module is designed to extract features from multiple receptive fields to enhance the information exchange in the convolutional layer.Finally,a pyramid multi-scale feature fusion module is designed to efficiently fuse feature information at different scales.(2)Aiming at the problem of high computational complexity of significance detection,a multi-scale saliency rapid detection method combining edge information is proposed.First,the object-level edge detection network is combined to design a pyramid feature enhancement module,which can effectively extract features at various scales and reduce computational complexity.Then,a semantic supervision module is designed to extract the global information by magnifying the receptive field at the top level of the network.Semantic supervision is performed on the upsampling process.Finally,a loss function that combines edge information is designed to efficiently learn the edges and salient regions of the model.(3)For JPEG compression applications,a JPEG image compression model based on salient regions is designed.Firstly,use different compression factors to compress and store the compression value for easy query;then normalize the salient image.Then,traverse the salient image and select the compression factor of the pixel according to the pixel value.Finally,query the compression value according to the compression factor.(4)For point cloud data processing applications,a saliency detection model for point cloud space is designed.First,the Kinect camera is used to complete the collection of color and depth images.Then,the saliency detection is performed on the color image to obtain a saliency binary map.Finally,whether to perform point cloud rendering is determined based on the binary map.