Study On An Image Saliency Detection Model Based On The Unified Bio-inspired And Statistical Analysis

Visual attention mechanism is an important reason of why the human locates regions of interest in a complex scene rapidly.Study on visual attention is originated from cognitive science and neurobiology,and now developing rapidly in computer science.Visual attention is quantificated by the term visual saliency in computer science,and visual saliency models analyze the visual attention mechanism and design mathematical models to simulate this mechanism.After more than 20 years of development,visual saliency detection models are approaching the habit of humans when processing visual information.But the concept visual saliency is ill-defined,visual saliency detection is still an open-ended question.The causes of this phenomenon lies in two points,firstly,studies on the mechanism of visual attention in cognitive science and neurobiology is imperfect,secondly,the definition of visual saliency in computer science needs improvement.Therefore,how to define visual saliency is the key to study visual attention in computer science.This paper investigates how to define visual saliency by two ideas.The first idea is simulate the selective mechanism of human visual neuron receptive field,and we propose a bio-inspired visual saliency detection model.The paper defines a template to simulate the receptive field of visual nueron,and convolves with the input image,the result predicts human eye fixations,this model called RFS.Furthermore,since the receptive fields of simple cells in the primary visual cortex produce a sparse representation of input signals,we train a non-salient template based on RFS,this template represents non-salient features of large amount of natural images.After computed the response of the input image and the template,we obtain regions that do not attract human visual attention,then we infer the salient regions inversely,this model called SRS.Both model RFS and SRS are bio-inpsired,experimental results show that they run fast,and the accuracy is higher than state-of-the-art models,and SRS performs even better than RFS.Because the former idea has not considered about the intrinsic correlation of image features,only the prior knowledge(i.e.,the selective mechanism of human visual neuron receprive field)is involved,the second idea is concentrate in the intrinsic features of the image.The second definition of this paper is based on statistical analysis,we analyze the similarity between outliers and human eye fixations in experiments,and propose an idea that measure saliency of a feature by its outlierness.In practice,we use basic distance based outlier detection to compute boolean maps,multi-scales of boolean maps are linear combined in order to form the saliency map,this model called OS.Furthermore,since the time complexity of basic distance based outlier detection model is O(n2),we use one time sampling approach to reduce the time complexity,the time complexity of the improved model is O(sn)(s is a constant value),this model called OSOS.The paper analyze the robustness of the algorithm,and prove that the robustness of the algorithm will not inflluenced by a small sample size.Both model OS and OSOS are based on statistical analysis,expeirmental results show that the accuracy of the proposed models are higher than state-of-the-art models,and the running time of OSOS is much shorter than OS.The two ideas employ two different definitions of visual saliency to detect image visual saliency.According to our experiments,the human eye fixation prediction results of the two saliency definitions are entirely different.This paper proposes a unified model which based on SRS and OSOS,the unified model utilize prediction error theory and guided filter,and proposes a new definition of visual saliency,this new model called SROD.The model SROD fuses the benefit of SRS and OSOS,and the expeirmental results show that SROD outperforms state-of-the-art models.Study on visual saliency consists of two major tasks,human eye fixation prediction and salient object detection,they are considered to be independent tasks in the past.This paper proposes a salient object detection model called OS2,it is translated from the former proposed model OS,the major modifications are image representation,feature representation and post processing,the saliency definition is also depend on the outlierness of the feature.Experimental results show that model OS2 outperforms state-of-the-art models,and it illustrates that it is possible to translate an eye fixation prediction model to a salient object detection model.