Salient Region Detection Based On Classification

In recent years,by the visual salient region detection,people learn the interests of the images,has become one of the hot areas of computer vision.The region of interest is a significant area,and the most effective way to identify the significant region is to establish a set of visual attention model.Through previous studies we can learn that the attention model can't work without the visual attention mechanism.While the visual attention mechanisms obtain two different ways:Bottom-up approach and top-down approach.The top-down approach is task-driven and is influenced by high-level semantic information;while the bottom-up approach is driven by data and is influenced by underlying information.The existing salient region detection methods usually use one visual attention mechanism,so that the results have difference with the true map.To solve the problem,this paper combines the bottom-up and top-down visual attention mechanism to establish a visual attention model,the main work includes:(1)Analyze the advantages and disadvantages of the classical Itti algorithm to extract features of the image,and propose a salient region detection of the images in combination with the bottom-up and top-down visual attention mechanisms.In the bottom-up visual attention,for the shortcomings of the original method,we use the Dense Sift method,which has not been applied to salient regional detection,to extract features of images,and then put the features obtained by these two methods to the classifier.Compare and analyze the experimental results,and verify the feasibility of using Dense Sift method in salient detection.In the top-down visual attention,we get the eye movement data of the image by performing the eye movement experiment,and the ground truth is obtained by Gaussian convolution.Then we put the positive and negative samples which are selected in the image to add to the late classifier for training a new model.(2)We proposed the DENSA model which is based on Dense Sift and Adaboost method to carry out salient region detection.This paper analyzes the four Adaboost methods of Adaboost,and finds out the best method for feature learning as the ultimate strong classifier.We put the positive and negative samples and the corresponding features obtained by eye movement experiment to the classifier for training,and get the final DENSA model.The comparative experiment with linear SVM and other comparative experiments of existing models confirm the advantages of DENS A model for salient region detection.(3)As a follow-up experiment of this paper,this paper attempts to use the deep learning method to extract the image convolution layer features for salient region detection,We get the features map of the original image size by up sampling the features of the neural network convolution layer,and then the subsequent correlation operation is carried out,and finally verify the feasibility of the method by the relevant experiments.