Research On Visual Attention Mechanism Based Saliency Objects Detection And Extraction Algorithms

In recent years,the Internet and computer technologies are flourish,image information increases in erupting trend.The information contained in the image is very rich,how to location interested objects and get the appropriate information quickly from a large number of information is the problem to be solved urgently.As a popular research field,computer vision is devoted to the rapid acquisition and application of object information from images or videos by simulating human visual mechanism.As the main source of visual information,image is the research emphasis in computer vision.The introduction of visual attention mechanism provides new ideas for the computer vision technology to position interested objects rapidly and to reduce the redundant information processing,which has aroused the study enthusiasm of scholars at home and abroad.The saliency is the expression of the visual attention mechanism in the computer field,in which the saliency object detection is an important research content.The core part of the saliency object detection is the study of the saliency calculation and extraction algorithms.Based on this,the saliency object detection and extraction algorithms are studied.The main research contents are as follows:(1)Propose a Otsu threshold segmentation method based on improved cuckoo Search algorithm(MSCS)to solve the problem that the traditional Otsu method has the effect of over-segmentation and is susceptible to noise,and the cuckoo algorithm(MCS)which introduces the stable random number mechanism can not solve the problem of low activity and slow convergence at later stage.Introduce a dynamic step size control factor to improve the ability of searching solution and accelerate the convergence.Introduce an inertia weight factor to reduce the local optimal situation and improve the precision of the solution.Combine the MSCS with the Otsu method to segment images,then compare and analyze the results effect with the similar algorithms.(2)Propose a multi-scale local-region saliency detection(LRS)method based on covariance to solve the problem that the local saliency detection may loss details of objects in spite of contour salient.Build a new model by analyzing the deficiency of existed saliency detection method.Introduce a regional detection method to solve the problem of detail loss on the basis of local saliency detection method based on covariance(CovSal).And extend the integrated local-region saliency to multi-scale by linear combination producing the final saliency map.Then,compare and analyze the detection results of proposed LRS with the similar algorithms.(3)Propose a saliency target extraction method combined with segmentation fusion strategy based on MSCS and LRS to solve the problem that it's difficult to segment targets accurately after a saliency detection sometimes which is based on linear fusion strategy.Firstly,extract the local and region saliency separately by the MSCS,merging them linearly in a certain proportion and extending the result to multiscale situation to extract the objects.Compare and evaluate the extraction results with similar methods under the standard data set and actual complicated scenes to prove its effectiveness.