Research On Color Image Segmentation Algorithm Of Semi-supervised Spectral Clustering Based On Firefly Optimization And Noticeable Difference

With the developing of multimedia imaging technology,color image has been the main medium in daily life.Segmentation and analysis of color images are always the hotspot problem in machine learning and computer vision,and have attracted wide attention of many researchers.Spectral clustering algorithm can cluster samples in arbitrarily distributed space and converge into the global optimal solution.It can achieve better results than traditional clustering algorithms in image segmentation.However,there are two problems when applying traditional spectral clustering algorithm to image segmentation.The first one is that this algorithm need to construct the similarity between arbitrary two pixels,which leads to the large similarity matrix and decompose problem on this matrix.Because of the large similarity matrix,there exits huge computation and storage problems.The other is that the traditional spectral clustering algorithm often uses the first k eigenvectors to reconstruct the sample space.It may lead to omit the other effective eigenvectors combinations.In order to solve the related problems of spectral clustering algorithm,this thesis proposes a semi-supervised spectral clustering color image segmentation method based on firefly algorithm and just noticeable difference.In the proposed methods,some supervisory information is extracted from the image,the just noticeable difference theory and firefly algorithm are introduced,and the key pixels are selected from the image by using just noticeable difference theory and firefly algorithm.After obtaining key pixels,the similarity measure method based on the connectivity and discreteness is constructed to obtain the effective segmentation result.In addition,the eigenvectors of spectral clustering are selected by using the firefly algorithm to ensure to be a more ideal data representation.Spectral clustering algorithm under this effective eigenvector combination can obtain satisfying results.The main work can be summarized as follows:(1)In the segmentation method based on just noticeable difference theory,the selection of just noticeable difference threshold is an urgent problem to be solved.To solve this problem,a JND color image segmentation based on firefly algorithm is proposed.Firstly,the supervisory information is adaptively extracted by using the neighborhood information of the pixels.Then the intelligent optimization algorithm based on firefly is used to select the threshold.Finally,image segmentation results are obtained.From the experimental results,we can know that the proposed algorithm outperforms the traditional just noticeable difference image segmentation algorithm in color image segmentation.(2)A semi-supervised spectral clustering algorithm based on firefly algorithm for color image segmentation is proposed to solve the problem that the performance of spectral clustering algorithm is easily affected by similarity measurement when applied to color image segmentation.The algorithm uses a small amount of supervisory information provided by users,utilizes the optimization method of firefly to optimize the threshold to obtain the key pixels of color image,and then constructs a similarity measure with connectivity and discreteness for these key pixels Based on this similarity measure,spectral clustering can cluster key pixels and the segmentation result of the image is obtained.The experimental results show that the proposed method is feasible and can obtain the ideal segmentation results.(3)In spectral clustering algorithms,the first k eigenvectors are often used to represent original samples.However,they may not produce satisfactory clustering results.Based on the consideration of eigenvectors selection,we propose a spectral clustering segmentation algorithm based on eigenvectors selection by firefly.In this algorithm,ordinal number of eigenvectors are encoded and the firefly algorithm is used to select the best combination of eigenvectors.Experiment results show that the proposed method presents effectiveness on image segmentation.