Research On Image Saliency Detection Method Based On The Adsorption Model

In the digital image process,the saliency detection is detecting the most important,attentive and informative part of a scene.Since the 1990 s,the saliency detection was first proposed up to now,it has developed into a very important branch in the field of computer vision analysis.In recent years,more and more people committed to the study of visual saliency detection,which show that the visual saliency detection has a great relationship with human visual attention system.In the human visual attention system,when people start to see the visual scene,the human will first attend the focus of attention.Then the information near the focus of attention will be processed further to obtain the saliency areas or objects.Then the other focus of attention will be searched in the scene,if there is a focus of attention,the visual attention mechanism will continuous deal with the information near the focus of attention.In this paper,the law of force in physics is used to mimic the visual attention system and an adsorption-based method which converts the saliency detection to a continuous adsorption process is proposed.The method includes the following steps:First of all,the input color image is converted into the CIELab and CIEHSV color space.The FCM clustering algorithm is used to cluster in the SHba L~*,a~*,b~*,H~*,S~*color channel,respectively.The aggregation degree of every cluster is used to identify the candidate salient regions and background regions.Next,the color differences,attention degree,and center offset degree of each candidate salient region are calculated,which combined with the aggregation degree are used to calculate the saliency of each candidate salient region which is used to obtain an saliency map.The Winner-Take-All network based on leaky integration-and-fire model in the saliency map is used to identify the focus of attention.And then,the salient regions are continuously adsorbed from the candidate salient regions.The specifice way is: For the current candidate salient reiogn to be judged,the adsorption forces which the candidate salient region receives from the focus of attention and the background regions,respectively,are used to determine whether the candidate salient region is salient..At last,in order to obtain a saliency map,the color difference and space distance are used to measure saliency of each pixel.Extensive experiments are performed on ASD database including 1000 natural image.And the results of saliency maps and performance comparison with the state-of-the-art algorithms demonstrate that our algorithm achieves better performance.