Research On Visual Saliency Detection Method Combining Time Domain And Frequency Domain

At present,environmental perception technology is one of the hot topics in the research of intelligent mobile robots.The environmental perception of mobile robots is a key issue in the development of robottechnology.The progress in environmental perception technology will lead to a dramatic development in the field of mobile robots.It is generally believed that visual saliency detection is a feasible way to solve the problem of robot environment perception.Image saliency detection introduces the visual attention mechanism into computer image analysis by simulating human visual perception process,enabling it to extract pivotal regions from environmental image data,reducing computational complexity of high-level environmental visual comprehension algorithms.Nowadays,saliency detection has been widely used in many practical visual applications such as image retrieval,image segmentation,video tracking,target recognition and robot navigation.Current saliency detection algorithms can be generally divided into two categories:saliency detection based on spatial domain and saliency detection based on frequency domain.However,no matter what the detection method is,in reality,due to the complex background of the image,the interference of the uncertain noise and the incompleteness of prior information,a single spatial domain detection or frequency domain detection method cannot effectively balance the needs of the algorithm complexity and accuracy of detection.This paper aims to improve the accuracy of saliency detectionwhile ensuring that the complexity of the algorithm.Firstly,a multichannel feature fusion algorithm is developed to integrate the color and depth information RGB-D data.Then,the hyper-complex Fourier transform is applied in frequency domain to get the multi-scale saliency maps.After that,an uneven super-pixel segmentation algorithm is used to smooth each obtained saliency map.In this way,the interference of discrete background noises is eliminated and the detection result in frequency domain is improved.Finally,the cellular automata algorithm is adopted to merge the multi-scale visual saliency maps and extract the final saliency region.The innovation of this paper and the main work is:(1)A saliency detection algorithm for joint space spectrum analysis is proposed.On the one hand,super-pixel segmentation of the image in the spatial domain is carried out while calculating saliencyin the frequency domain,helping increasing the consistency of saliency detection in local regions and improving the fusion effect based on cellular automaton.On the other hand,in order to combine the natural fusion of the depth channel features of RGB-D data and the RGB color channel features,a quaternionis constructed by using super complex numbers,and the fast convergence of the feature level in the frequency domainis integrated.(2)An image visual distance calculation methodbasedon saliency detection is used.In the process of calculating the image distance,the valuable information in the image can be highlighted more effectively and the matching accuracy is improvedby considering theprior information of saliency,adjusting the distance-to-weight ratio,highlighting the weight of saliency regions,and suppressing the background region.Finally,the histogram matching algorithm is used to achieve the final image matching.(3)A visual saliency detection method based on spatial-spectral mixture analysis is applied to the robot vision perception.The algorithm can quickly and efficiently extract the visual saliency regions in complex real scenes,and achieves good balance between the algorithm efficiency and the detection accuracy,which makes the saliency detection serve themobile robot of the visual detection system better,and enhances the perception performanceofthe surrounding environmentof the robot in real scenes.