| With the development of computer vision technology and the large-scale use in production and life,image saliency target detection is becoming more important as an image pre-processing process.It can detect areas noticed by the human eye in the image,and use the useful information in the picture.The screening of information not only reduces the amount of calculation,but also increases the calculation speed,greatly saving people’s time and improving people’s work efficiency.In this paper,the image saliency target detection is mainly studied from three aspects: the fusion of image discriminative region features and label propagation,the fusion of deep features with traditional diffusion random walks,and the enhancement of local features of neural networks.First,in order to improve the accuracy and robustness of image detection in complex scenes,this paper proposes a saliency target detection algorithm that fuse discriminate region features and label propagation.Use the distinguishing region feature method to form a distinguishing region feature saliency map;at the same time,use the label propagation saliency map obtained by the label propagation algorithm to optimize the distinguishing region feature map;combine the two by an exponential function to form the final target saliency map.This method makes the target area more prominent,and the background area is more suppressed.Secondly,in order to solve the problems of incomplete feature learning of traditional methods in image saliency detection and salient area salientness in complex scenes,this paper proposes a saliency detection method based on multi-level depth features and random walk.Use the full convolutional neural network to combine the deep and shallow feature information to extract the multi-level convolution features of the picture;segment the picture by super pixels,assign the extracted deep convolution features to the corresponding super pixels,and construct the feature matrix;The eigenmatrix of the model uses regularized random walk ranking model to generate the final saliency map.This algorithm combines the advantages of convolutional neural network saliency detection method and traditional saliency detection method,does not require large-scale model training,and also achieves a higher accuracy than traditional methods.Finally,in order to improve the utilization of neural network features,this paper proposes a saliency target detection algorithm for local feature enhancement learning.In this paper,by combining local and global information,the multi-scale context feature extraction module is used to effectively integrate multi-scale features to more completely extract picture details and make the saliency map more accurate. |
