Research On Bottom-Up Visual Saliency Detection Model

The purpose of saliency detection is to find the most attractive areas of the image.With the rapid development of computer vision,saliency detection has become a popular research direction.The main task of saliency detection is to make the computer can simulate the human visual system attention mechanism,which can quickly find the target area in the image.At present,although scholars have proposed a large number of methods for saliency detection from different perspectives,these methods still have problems such as insufficient target area and poor suppression of background noise.In view of these deficiencies,the thesis proposes two kinds of saliency detection algorithm.The main research contents are as follows:(1)We propose a multi-feature optimal fusion saliency detection algorithm.It is important to choose the feature that can represent the saliency of the target area of the image.Through our analysis,we finally select the color features,texture features and frequency domain features of the image to detect the saliency area of the image.Compared with the previous saliency detection algorithms which only add or multiply image features,the saliency detection algorithm based on multi-feature optimal fusion proposed can learn the training data by Adaboost.Different weights are set for various features according to their different contribution to saliency detection,which can make the results of the saliency detection more accurate.The experimental results show that the P-R curve of this algorithm is better than most algorithms,and other performance indicators also have good performance.We can get a more accurate target area.(2)We propose a new saliency detection algorithm based on multiple priors and global contrast.Because the use of the priori information of the image can help distinguish between target area and background.Therefore,the thesis combines the background prior information,the central prior information,the foreground prior information of the image and the global color contrast to detect the saliency area.Combine the advantages of these three prior information with the global color contrast can improve the effect and help filter the background noise.At the same time,this paper uses a cost function to optimize the saliency map.The main idea is to design a saliency map cost function by analyzing the contrast relationship among the foreground,the background,the target area and the role of the central priori in the saliency detection.The saliency value of each super-pixel is obtained by solving the optimal solution of the cost.The experimental results show that compared with the classical algorithms,the proposed algorithm has a good improvement on the P-R curve and the MAE value.At the same time,it can obtain a more accurate target area with clear edges.(3)We propose a saliency detection algorithm based on semantic enhanced convolutional neural network.The model is improved by the classic model VGG16,in which the fully connected layers are replaced by the convolution layer.And the BN operation is added to the convolution to speed up the training of the network.The Dropout operation is added to the convolution layer.And SENet module is embedded behind the final convolutional layer to enhance the semantics of the extracted features.The method can extract the features of the deep layers of the image and adaptively weight the extracted features.Compared with other algorithms,the algorithm has some advantages in subjective performance and objective indicators.