Visual Saliency Detection Based On Multiple Feature Fusion

Facing massive information from visual input,the human visual system quickly finds out the interested objects by selective attention mechanism and prioritize the processing of important visual information.With the image acquisition technology,Internet communication technology and multimedia processing technology have infiltrated into people's daily work,study and life,and the rapid growth of image data,the problem of information redundancy is coming.It is very important to find a suitable computational model for simulating the human visual attention mechanism makes computers have human-like capabilities of information processing,and this resulted in visual saliency detection.Visual saliency detection can extract the salient object from images which provide important reference information for image analysis and image processing.The results of visual saliency detection have been widely used in the applications of pattern recognition and computer vision such as object recognition,object segmentation,image compression and image retrieval.Currently,the research of visual saliency detection has attracted much attention.Aiming at the problem of existing visual saliency detection methods mainly focus on the underlying feature of images,the fusion mechanism is single,and the accuracy and clarity of the detected saliency regions are not high.Based on the selective attention mechanism of human visual system,combining image pre-segmentation,color space,color variance statistics,underlying features,high-level prior information,fusing features in multi-channels,after using multiple fusion methods to fuse feature maps,the saliency map was extracted.The main work is as follows:(1)In the existing visual saliency detection methods,the fusion method is simple,the high-level prior information is ignored,and the accuracy of the saliency detection is not high.A visual saliency detection based on fusion of color and texture features(VFCT)is proposed.On the basis of color space conversion and SLIC superpixel segmentation,the color feature map is extracted by fusing color contrast feature and color distribution feature,and the texture feature map is extracted from the texture feature.According to the texture complexity of the image,the initial saliency map is obtained by fusion of color feature map and texture feature map.The final saliency map is obtained by integrating the center prior information to the initial saliency map.The experimental results show that this method can effectively improve the accuracy and clarity of saliency detection,and the salient object has better integrity and clarity.(2)In order to further filter the background noise of the image,increase the degree of separation between the salient object and the background region,combining the characteristics of human visual system,a visual saliency detection based on visual attention point prediction(VAPP)is proposed.After color space conversion and SLIC superpixel segmentation of the image,color variance statistics are performed on multiple channels of multi-color space to complete visual attention point prediction,and the initial saliency center of the image is obtained.Based on the initial saliency center of the image,the method of extracting color contrast feature and color distribution feature is improved,and the two features are fused to get a new color feature map.The initial saliency map is obtained by fusion of new color feature and texture feature.The final saliency map is obtained by incorporating the location feature map which extracted based on the initial saliency center into the initial saliency map.The experimental results show that this method performs well in the filtering of background noise can obtain a more complete and reliable salient object,and the detection performance is obviously improved.