Consistent Image Processing Based On Co-saliency

In recent years,many achievements have been made in various fields of computer vision,but there are still gaps in the field of image saliency redirection.The image saliency redirection algorithm is dedicated to processing the local features in the image,and achieves the purpose of enhancing or weakening the saliency of the region so as to guide the user's visual attention.It can be used in image editing,smart advertising,and improving image aesthetics.In order to achieve the effect of guiding the user's visual attention,we first need to know which information is more interesting to the human eye,so we introduce an image saliency detection algorithm.Image saliency detection algorithms have achieved many results in computer vision related fields,among which joint saliency detection algorithms are widely used in image co-segmentation,target detection and other fields.The main research content of this paper is the image saliency redirection in the image group.The purpose is to redirect the non-salience objects in the image group to reduce their saliency,so that the co-saliency objects in the image group are in each image.There is a unique saliency,which achieves the effect that the single-image saliency map after the redirection process is basically the same as the co-saliency map.We call it "consistent image processing".In order to find the non-joint salient object area more accurately,this paper proposes a new image joint saliency detection algorithm and introduces matting technology to extract the non-joint salient object area,and finally performs image saliency redirection in HSI color space modeling operating.The innovation points are as follows:1.This paper proposes a co-saliency detection algorithm based on deep learning.First,the residual network ResNet34 is used as a feature extraction network to obtain image features at different levels and the common convex feature matrix of the image group,and calculate the correlation coefficient of the image group Matrix.Second,using gradient back propagation for deep feature induction and screening key features for weight distribution.In order to ensure that the features are not diluted when passed layer by layer,we added the attention mechanism CBAM(Convolutional Block Attention Module),and finally passed the top-down in the decoding stage The iterative calculation of to obtain the final co-saliency map.2.This paper proposes a KNNT matting algorithm based on enhanced optimization,mainly for the previous matting algorithm to directly operate on the original data set,which makes the matting results obtained on low-contrast images and images with similar colors of the front background not Ideal question.First,the optimized Gaussian high-pass filter is used as the homomorphic filter function,and the image is regarded as the multiplicative combination of the illuminance component and the reflection component,and then the input RGB image is subjected to homomorphic filtering in the three channels.This operation is enhanced Contrast is used to improve the sharpness of the image,and the Gaussian highpass filter can enhance the details and edge information of the image.Secondly,in order to avoid homomorphic filtering amplifying noise interference while enhancing high-frequency signals,this paper will perform a contrast-limited adaptive histogram equalization on the filtering result in the HSV color space,and separately process the saturation component S and the brightness component V,hue The component H remains unchanged.Then,the image texture features extracted from the enhanced image are richer,and then the texture features are logarithmically transformed to compress the range of feature values.Finally,the texture feature is introduced into KNN matting,and the final mask result is obtained through closed solution.3.Aiming at the interference problem of the non-joint salient objects in the image group,this paper proposes an image saliency redirection algorithm,which can guide the user's attention by modifying the local saliency in the image.After using the joint saliency detection algorithm and the natural image matting algorithm to obtain the precise position information of the non-joint salient objects,we model the hue,saturation,and brightness in the HSI color space.In terms of hue,the hue distribution is modeled,and it is believed that when the non-jointly salient object area and the surrounding background area have the same or similar hue,the optimal hue solution is obtained.On the two channels of brightness and saturation,we allow users to set the parameter saturation difference and brightness difference according to their own wishes.When the saturation and brightness model reaches the entire set difference value,the optimal solution is obtained.The experimental simulation of the algorithm and the analysis of objective evaluation indicators show that the algorithm proposed in this paper is feasible and effective.