Research On Saliency Detection Models Of Stereo Image

According to the human visual attention mechanism,people will always pay attention to some objects or regions that are prominent relative to the background in a scene,and ignore some parts that are uninterested to reduce the complexity of visual processing.Saliency detection is to simulate the mechanism of human eyes and extract the region of interest from the image.As an image preprocessing process,saliency detection can be used for many visual tasks,such as image quality assessment,visual tracking,target relocation,image classification,image segmentation and so on.With the development of image acquisition technology,more comprehensive information such as depth prompt,correspondence between images or time relationship can extend 2D saliency detection to stereo image saliency detection.How to effectively use depth information to build a depth-induced saliency detection model and how to realize the best combination of2 D view and depth view is an challenge for the detection of stereoscopic image saliency.This paper first explored the method of building saliency detection model by manual feature extraction,then proposed two optimization models of stereo image saliency detection combined with cellular automata,and finally explored the application of stereo image saliency in image quality assessment.The main contents are as follows:(1)A saliency detection model based on band-pass filtering and saliency map fusion optimization was proposed for 3D images,which is detected by combining texture and depth features with image saliency detection.Firstly,the saliency map of the left image are calculated by the improved graph-based manifold ranking algorithm;Secondly,we extract texture feature from left image to compute texture saliency map with a Log-Gabor filtering method,and extract depth feature from stereo image to compute depth saliency map with a Log-Gabor filtering;Thirdly,the three saliency maps are integrated into a stereo saliency map by the weighted linear combination(WLC)method;Finally,the 3D saliency map is enhanced by the center-bias factor and visual acuity.Experimental results on a public eye tracking database show that the proposed model achieves better detection performances than the existing 3D visual saliency detection models.(2)We designed two kinds stereo image saliency detection optimization method based on cellular automata,and compared the performance of them.The strategy of the first method is to firstly enhance the 2D saliency map and the depth saliency map with single-layer cellular automata,and then merge them with the texture saliency map;In the second method,2D saliency map,depth saliency map and texture saliency map are directly processed by multi-layer cellular automata.Multi-layer cellular automata can fuses multiple saliency mapswhile integrating the advantages of each saliency map.Finally,the obtained stereo saliency map is optimized by using the method of center bias and visual acuity.(3)A three-branch convolutional neural network model based on the convolution network was constructed to predict the quality of stereo images.We took the normalized stereo image pair and the stereo saliency as the input of the model.The left and right view branches have the same weight based on Siamese network principle,and the comparison loss function is used to optimize the weight value of the network,while the branch of stereo saliency map learns the data characteristics independently.Finally,the feature vectors outputted by the three branch networks are concated and connnected to the full-connection layer for regression.The last full-connection layer outputs the predicted score of the stereo image.Experiments on LIVE 3D Phase I database and LIVE 3D Phase II database show that the proposed method has good prediction results.