Visual Saliency Detection Framework Based On Sparse Coding

With the image and video data scale increasing more and more rapidly,the lack of efficiency in the algorithm becomes the main issue.Whereas,researches in saliency catch its significant popularity.In this paper,we proposed a general visual saliency model based on sparse coding,which takes into account the bottom-up and top-down stages of visual attention.According to characteristics of the sparse coding as well as the discrete expression in visual signals,the saliency in saliency signals is related to sparse expression,which we demonstrate in the paper,the elaboration of the relationship between sparse coding and visual saliency.In our method,three Hypotheses are put forward.Firstly,the base in our brain is over-complete,basing on which encoding objects will generate the new base space,in short,a high-level dictionary.Secondly,the brain tends to encode with the active high-level dictionary.Lastly,saliency is determined by the response to the dictionary.Hence,as we propose,in the bottom-up stage,the high-level dictionary is determined by the peripheral signals of the candidates,while in the top-down stage,it depends on the top-down cues such as the task,priori knowledge,personal interests,and so on.The main contribution in our paper lies in the fact that we introduce the training of discriminant dictionaries to establish a visual significant model framework,which breaks in the barriers across the bottom-up and top-down method.The experiments of general open-access bottom-up test set together with the aerial remote sensing image dataset show that the model has achieved excellent results in the two stages of the visual attention respectively.Compared with other similar saliency models,our proposed method outperforms in the following aspects.First,it based on the cognitive psychology explanation.Second,both the bottom-up and top-down stages can be abstracted as a unified description using the same interpretation.Third,with our simple and easy-to-use algorithm,it can be widely applied to target recognition,segmentation and other image processing.