Research On Visual Saliency Combining Spatial Consistency With Multilayer Structure Information

In recent years,visual saliency detection has been one of the hot field of computer vision.It extracted the information from the image region of interest,which ensures the human eye to obtain and process image information efficiently.Currently,visual saliency detection model has been widely used in many fields,such as object recognition and detection,image quality assessment and image retrieval.It is full of challenge due to the complexity of the object structure,the background interference and so on.This paper is mainly to study how to combine the spatial consistency and multilayer structure information to improve the accuracy of visual saliency detection.This paper first studies the application of low-rank matrix representation method in the field of computer vision,analyzing its advantage of feature expression and image representation,then proposes a top-down saliency detection model based on structure low-rank coding(SLRC).Based on low-rank representation and joining the dictionary learning mechanism,this model can effectively combine spatial consistency and structure information consistency and extract different information between image blocks,which provides a more discriminating image block representation for visual saliency detection.Conducting experiments on Internet car datasets and Graz-02 datasets containing three subsets,which are person,bicycle and car,the average detection accuracy of the model proposed in this paper are raised by 3.3%,1.6%,5.6% and 6.7% respectively compared to the saliency detection model based on classification framework.Experimental results show that this model can effectively combine spatial consistency with structure information consistency and require more discriminate representation of image blocks,which can enhance the detection performance of salient object.This paper also studies the local feature pooling method in saliency detection and proposes a multiple sub-block context spatial pooling method(MSB-CSP).This method can extract multilayer structure information,which is the neighborhood context structure information and the context structure information within different spatial direction.Meanwhile,this method enhances the robustness to scale changes by adding an information filtering function.Based on this method,combining with structural low rank coding,this paper further proposes a saliency detection model utilizing multilayer structure information,which can effectively combine spatial consistency with multilayer structure information,providing a more discriminating representation of image blocks.Conducting experiment on three target object subsets in Graz-02 datasets,compared to the context spatial pooling method,the average accuracy of the proposed MSB-CSP method are increased by 3.8%,3.9% and 5.1% respectively and the average accuracy of saliency detection model utilizing multilayer structure information are increased by 7.9%,6.4% and 11% respectively.Experimental results show that the proposed MSB-CSP method and saliency detection model can effectively improve the average accuracy of visual saliency detection.Meanwhile,when the scale size continues to increase starting from the optimal scale,the average detection accuracy of the MSB-CSP method is essentially unchanged,which illustrates that this method has strong robustness to the change of scale.