Research On Image Saliency Detection Based On Parallel Convolution Neural Network

Image saliency detection is to detect the area of human eye in the image.It has a wide range of applications in image processing,such as image tag,semantic segmentation,image compression and image content perception.Efficient image saliency detection algorithm can extract valuable image regions,greatly improving the efficiency of image processing.In recent years,as a key technology in computer vision research,image saliency detection has attracted a lot of domestic and foreign scholars' attention.In the current image saliency detection algorithm,the traditional method of manually extracting features based on mathematical theory to obtain the image saliency region can detect the details of the saliency object well,but there is a lack of global information in the detection of the saliency object,and it is relatively sensitive to noise.The detection algorithm based on semantic features can accurately detect the representative categories Significant objects,but the ability to detect the details of significant objects is slightly inadequate.For this reason,this thesis focuses on the research of image saliency detection based on parallel network structure.There are three main studies included in this dissertation:(1)Aiming at the situation that single type of saliency information can't solve the location of saliency object and the boundary between saliency object and background well,a method of detecting image saliency by extracting mixed features is proposed.This method integrates the contrast information of image region and the semantic information of image.After fusion,the mixed feature map has the ability to deal with the problem of salient object location and the problem of thinning the boundary between salient object and background at the same time.(2)In the process of constructing mixed features,a method based on region contrast information mapping map is proposed.This method extracts color features,texture features and spatial position features of each pixel cluster,calculates the differences between the target pixel cluster and other regions in the image,and generates region contrast information map with the same size as the original imageMapping.It effectively solves the problem of size mismatch in the input of convolutional neural network.(3)In order to solve the problem that a single network structure can not deal with the feature information of multi class salient objects,a method of image significance detection based on parallel convolution neural network structure is proposed.The region contrast feature map network is constructed to process the region contrast information,and the classical network model is used to process the image semantic information.At the end of the parallel network structure,two highly abstract feature maps are fused and the final significant map is obtained by classifier.It enhances the ability of the network to analyze the features of multi class salient objects,and makes the final saliency map more accurate.