Saliency Detection Based On Spatial Domain And Frequency Domain

Selective visual attention possesses an ability to rapidly and accurately find objects of interest in complex visual scenes,and it can reduce the occupation of neural resources.Saliency detection algorithm is to rapidly find salient objects in images by simulating visual attention.This kind of algorithm can effectively solve the analysis and processing of massive image data,eliminate non-important information interference,and reduce the complexity of computer understanding to image content.According to different attention mechanism,saliency detection algorithm can be divided into bottom-up detection method and top-down detection method.In this paper,bottom-up detection method is focused on,existing algorithms are learned to realize the shortcomings of different types of algorithms,and three novel saliency detection algorithms are proposed.The main research work of this paper can be summarized as follows:1.A saliency detection method based on visual center is proposed.Firstly,the input image is segmented into superpixels by SLIC algorithm,and the intensity feature channel L,color feature channel a and b,and coordinate position information x and y of the superpixel seeds points are extracted.Next,the difference saliency image is calculated by the difference values of the color and brightness feature channels.Then,the "visual center" is obtained through the difference saliency image,and the spatial weight of each super pixel is calculated.Finally,the final saliency map is generated by balancing the difference significance of spatial weights.Experiments show that this method can effectively suppress background interference and highlight salient targets.2.A saliency detection method based on hypercomplex domain is proposed.Firstly.quaternion hypercomplex is constructed based on intensity feature channel I and color-opponency feature channel RG and BY.Next,amplitude spectrum and phase spectrum is extracted.Then multi-scale wavelet transform is carried out on the amplitude spectrum to generate multi-scale visual saliency image.Finally,the evaluation function of transverse comparison is constructed and the final saliency map is generated by fusing the better saliency image according to the evaluation function.Experimental results show that this method can effectively suppress background interference,and the salient target is more complete.3.A saliency detection method based on discrete cosine transform is proposed.Firstly,respectively performs discrete cosine transform on intensity feature channel I and generalized R,G and B color feature channels,and extracts amplitude matrix and symbol matrix.Then,multi-scale wavelet transform is performed on the amplitude matrix of all feature channels to construct multi-scale channel saliency image,and scale saliency image is generated with the feature channel that the coefficient is weight factor.Finally,the evaluation function is used to select the better saliency image for fusion,and the final saliency map is obtained after the center-bias optimization.Experimental results show that the saliency map obtained by this method accords with the result of visual attention,and it has high accuracy.