Research On Bottom-up Visual Attention

Bottom-up visual attention is a kind of inherent cognitive ability with which human beings are able to find objects or regions that are clearly different from the surrounding area quickly and ignore other unimportant regions meanwhile.It can be applied in various fields such as extracting the object of interest,image compression,image retrieval,extraction of key-frame.It has significant value in research and practical applicationsAn image-oriented bottom-up visual attention calculation method was present in this paper.This algorithm contains four modules:extraction of early visual features,saliency computation based on the local spatial sparse coding and reconstruction,features competition and mergence,morphology operation.In the module of extraction of early visual features,we separated the information of three channels from the original image and reconstructed it into two color features,RG(Red-Green)and BY(Blue-Yellow)by "color double-opponent" mechanism.In order to drive the computation of bottom-up visual attention with the sufficient low-level information of the image we have obtained,after including the intensity as the third feature,we made an expression of all these three feature maps on multi-scale.In the module of the saliency computation,we attempted to use sparse coding to express the statistical significance of local regions,which corresponds to the principle of visual local simulation on human Retina.For the purpose of reducing information loss and showing the differences between local regions adequately,Locality-constrained Linear Coding(LLC)was adopted.This coding method takes the central region as input signal,and surrounding region as dictionary.The reconstruction error of coding represents the degree of deviation between central region and surrounding region,and it can be treated as saliency value of central region efficiently.In the module of features competition and mergence,we proposed a winner-take-all elimination strategy based on entropy,which only accumulate the result of representative features.Then,the mentioned image algorithm was verified on the MSRA-1000 Database.The experimental results showed that this algorithm has better performance comparing with the other nine bottom-up visual attention algorithms.Subsequently,we introduced cross-entropy evaluation method which was designed directly for the saliency map,and the algorithm we proposed got top grades,both on common evaluations such as Fi measure and ROC(Receiver Operating Characteristic).Then we made an extension on the basis of our image algorithm,and proposed a video-oriented bottom-up visual attention algorithm.This algorithm mainly contains three modules:extraction of early visual features,saliency computation based on the multi-temporal scale sparse coding and reconstruction,feature mergence and normalization.In order to reflect the difference between preceding and subsequent frames,and calculate the motion information of the video adequately,we proposed different kinds of coding and reconstruction strategy with multi-temporal scales according to different types of video.Meanwhile,a video dataset was built in our work.This dataset contains 16 clips of video in various types and topics.And the dataset was utilized to verify the feasibility and validity of this video algorithm.The experiment results showed that the video algorithm we proposed is able to obtain the most attractive regions in the video frames.And the performance of the video algorithm is better than other three algorithms in the evaluation of PRI(Probabilistic Rand Index)and consistency measure.