Research On Adaptive Spectral Clustering Algorithm And Application Of Flame Segmentation

As an effective clustering analysis method,spectral clustering algorithm is reliable and has good clustering performance.It has been successfully applied in various fields,such as text analysis,speech analysis,machine vision and image segmentation.Although spectral clustering algorithm has many advantages,it lacks self adaptability.It needs to manually set the scale parameter and cluster number of similar matrix,which is limited in practical applications.In this paper,we give an improved spectral clustering for solve the problems and shortcomings,and the improved algorithm is applied to the image segmentation field.The specific research contents are as follows:(1)For the traditional spectral clustering algorithm,we need to manually set up the scale parameter when calculating similar matrix.We propose a spectral clustering algorithm based on the natural nearest neighbor.First introduced the nature nearest neighbor search algorithm,and the termination condition of the search algorithm was improved to reduce the time complexity of the algorithm,and then use the density information to calculate the scale parameter for each data point,and perform the Gauss kernel similarity matrix calculation,finally,complete the clustering.Experimental tests on synthetic datasets and common test sets show that the algorithm improves the Fmeasure index by about 5%compared with the original spectral clustering algorithm and several typical algorithms.(2)A spectral clustering algorithm based on Eigengap is proposed to solve the problem that the number of clustering algorithms needs to be artificially set.By analyzing the relationship between the eigenvalues and the number of clusters of normalized Laplasse matrix,the concept of Eigengap is introduced,and the eigenvalue gap sequence is calculated,and the first maximum point of the sequence is found.The corresponding lower mark of this sequence is the number of clusters.Experiments on synthetic datasets and common test sets show that the algorithm can accurately determine the number of clusters,the fmeasure index of clustering results was increased by about 6%,and robustness of clustering results are high.(3)Aiming at the traditional K-means algorithm using random initialization method,the initial clustering center leads to unstable clustering results and easy to fall into local optimum.A spectral clustering algorithm based on density initialization is proposed.The density information is obtained from the nature nearest neighbor,and the points in the high density region is selected as alternative initial cluster centers,using the maximum and minimum distance method to get the final K initial clustering center of K-means algorithm,The emergence of unstable clustering results and local optima is avoided.This paper tests the performance of the algorithm in an common test dataset.(4)The improved adaptive spectral clustering algorithm is applied to the flame image segmentation.In this paper,an improved spectral clustering algorithm based on natural nearest neighbor and Eigengap is used in image segmentation.First of all,at the University of Berkeley standard library on segmentation test results show that this algorithm is better than the original clustering segmentation algorithm in accuracy,the segmentation results were increased by 4%to 7%on the RI index;then,the algorithm is used to flame image segmentation,on the Internet and share their own shooting flame image experiment results show that this algorithm has better segmentation effect.