Image Saliency Detection Via Laplacian Pyramid

People can get a lot of information that words cannot express from the images.As an important source of information,the study of image processing is getting more and more attention.With the rapid development of information technology,the scale and complexity of image is growing at the speed of the explosive,which has brought the enormous difficulty for subsequent analysis of the image processing.Saliency detection aims at highlighting salient foreground objects automatically from the background,and is an important and fundamental concept in neuroscience and psychology to investigate the mechanism of human visual attention systems.Firstly,the thesis fully analyzes several state-of-the-art saliency methods,then we found that the result of these existing methods can not produce better performance,while it only considers color feature or intensity feature.A single feature can be seldom well describe salient regions,because a single feature descriptor usually only captures one aspect of the visual information.For instance,the color-based descriptors may not handle the images with rich textures very well.To solve the problem,we also incorporate the color,intensity,and position information into the procedure and analyze various characteristics independently.Majority of saliency models consider visual features such as orientation,color,intensity and the importance of visual features has not yet been fully explored.Here,we take texture feature into consideration in particular,using the gray level co-occurrence matrix(GLCM)as a texture descriptor.On the basis of gray level co-occurrence matrix,the thesis propose a method of color co-occurrence matrix,which can reflect the information of color images in the direction,interval,amplitude and speed according to the color and texture feature between two pixels.In this paper,a simple method is proposed for detecting salient regions by utilizing muti-scale features at a superpixel-based level.Most saliency methods calculate salient value in pixel-grid manner,which generate saliency maps with low resolution,poorly-defined boundary.To address the above mentioned issues,we make full use of superpixel method by segmenting an input image into multiple uniform and compact regions,which can both improve the speed and improve the quality of the results.Meanwhile,this thesis presents a new approach based on Laplacian pyramidal decomposition to extract salient object.This thesis iterative images by using the method of Multi-resolution analysis based on Laplacian pyramidal decomposition and completes the multiscale decomposition.The important features and details of the image is decomposed to different sub image according to the characteristics of different scale.Then the corresponding layer images using different fusion operators are reconstructed to get the final fusion images.We evaluate the performance of saliency detection algorithms both qualitatively and by comparison to human observers.