A multi-objective optimization segmentation method for tree image based on fusion clustering and classification algorithm
-
摘要:
目的结合树木图像颜色和纹理特征，融合聚类和分类算法对树木图像进行多目标优化分割，从而提高自然背景下树木图像分割的准确性。 方法首先，利用MSCC框架理论，解决聚类和分类目标函数同时依赖于聚类中心的问题。然后，分别选定聚类性能评价指标函数和分类性能评价指标函数。最后，采用多目标进化优化方法——NSGA-II算法进行优化，得到Pareto前端最优解集，并通过计算聚类有效性指数I的最大值，寻找最优解决方案。选择具有代表性的法国梧桐、侧柏、松树和杏树等自然背景下拍摄的4幅图像作为样本。分别采用K-means、Fuzzy C-means、对聚类目标函数进行单目标优化，采用MOPSO方法进行多目标优化，以及NSGA-II方法进行多目标优化等5种方法对样本图像进行分割比较。 结果在聚类中心数量相同、种群大小相同、遗传代数相同的条件下，指数I的值表明本文提出的分割方法优势显著。对于4类不同样本图像分割的指数I值进行对比可知，以HF指数为单目标函数进行遗传优化的结果优于单一使用K-means和FCM算法；MOPSO多目标优化方法分割结果优于单目标优化结果；基于NSGA-II优化的多目标函数分割结果又优于MOPSO多目标优化结果。 结论融合聚类和分类算法构建聚类性能评价指标函数和分类评价性能指标函数，并采用非支配排序遗传算法对多目标函数进行优化，能更好地保留树木图像的颜色和纹理特征，分割准确率显著提高。 Abstract:ObjectiveIn order to improve the accuracy of tree image segmentation under natural background, this paper studies how to combine the color and texture features of tree image, and combine clustering and classification algorithm to optimize multi-objective segmentation of tree image. MethodBased on the tree image feature analysis, this paper proposes a multi-objective tree image segmentation method based on clustering and classification algorithm. Firstly, using the MSCC framework theory, the clustering and classification objective function depends on clustering center simultaneously. Then, the cluster performance evaluation index function and the classification performance evaluation index function were selected. Finally, the multi-objective evolutionary optimization method, NSGA-II algorithm was used to optimize, and the Pareto front-end optimal solution set was obtained. The I index was used to select the optimal solution from the optimal solution set. In this paper, we selected four images taken under the natural background, such as Oriental plane, Platycladus orientalis, pine and apricot, as samples. K-means, Fuzzy C-means, single-objective optimization of clustering objective function, multi-objective optimization using MOPSO method and multi-objective optimization using NSGA-II method were used to segment the sample images. ResultWhen the number of cluster centers, the size of population and the number of genetic iterations were the same, the value of index I can verify that the proposed segmentation method had significant advantages. Comparing the index I values of four different sample image segmentation, we can see that the result of genetic optimization using HF index as single objective function was better than that using K-means and FCM algorithm alone. The result of MOPSO multi-objective optimization method was better than that of single objective optimization method, but the result of multi-objective function segmentation based on NSGA-II optimization was better than that of MOPSO objective optimization results. ConclusionThe experimental results show that the segmentation accuracy of the method proposed in this paper is obviously better than that of single-objective optimization segmentation and K-means, Fuzzy c-means and other segmentation methods, the color and texture features of the tree image are better preserved. So the accuracy of segmentation is significantly improved. -
Key words:
- multiple-objective optimization /
- tree image /
- NSGA-II
-
图 1 两个目标函数的帕累托前端最优解集
Figure 1. Pareto front of a set of solutions in a two objective space
图 4 不同分割方法分割结果图
Figure 4. Comparison of segmentation results by different segmentation methods
表 1 NSGA-II实验参数设置
Table 1. Parameter settings for the experiment
参数名称Parameter 参数值Setting 种群大小Population size 20 遗传代数Number of generations 40 选择算子Selection operator 二进制锦标赛Binary tournament 交叉算子Crossover operator SBX算子SBX operator 变异算子Mutation operator 多项式变异算子Polynomial mutation operator 交叉概率Probability of crossover 0.8 变异概率Probability of mutation 1/染色体长度1/length of chromosome 交叉索引值Crossover index value 2.0 变异索引值Mutation index value 5.0 
下载: 导出CSV
表 2 不同分割方法指数I值比较
Table 2. Comparison of I index value by different segmentation methods
样本Sample K-means (K=3) FCM (m=2, K=3) HF指数单目标GA优化HF index single objective optimization MSCC多目标粒子群优化Two objective MOPSO optimization MSCC多目标NSGA-II优化Two objective NSGA-II optimization 法国梧桐Platanus orientalis 67.56 74.87 85.32 91.36 96.28 侧柏Platycladus orientalis 50.56 67.45 74.65 85.43 90.32 松树Pinus sp. 40.56 41.45 50.32 72.39 83.87 杏树Armeniaca sp. 59.67 65.34 69.56 78.82 87.98
下载: 导出CSV
-
[1] 葛玉峰, 周宏平, 郑加强, 等.基于相对色彩因子的树木图像分割算法[J].南京林业大学学报(自然科学版), 2004, 28(4):19-22. doi: 10.3969/j.issn.1000-2006.2004.04.004Ge Y F, Zhou H P, Zheng J Q, et al. Tree image segmentation algorithm based on relative color factor[J].Journal of Nanjing Forestry University(Natural Science Edition), 2004, 28(4):19-22. doi: 10.3969/j.issn.1000-2006.2004.04.004 [2] 赵茂程, 郑加强, 林小静, 等.基于分形理论的树木图像分割方法[J].农业机械学报, 2004, 35(2):72-75. doi: 10.3969/j.issn.1000-1298.2004.02.021Zhao M C, Zheng J Q, Lin X J, et al. Tree image segmentation based on fractal theory[J].Transactions of the Chinese Society of Agricultural Machinery, 2004, 35(2):72-75. doi: 10.3969/j.issn.1000-1298.2004.02.021 [3] 蔡世捷.基于Matlab的树木图像分割方法研究[D].南京: 南京林业大学, 2005. http://cdmd.cnki.com.cn/Article/CDMD-10298-2005103040.htmCai S J. Research on tree image segmentation based on matlab[D]. Nanjing: Nanjing Forestry University, 2005. http://cdmd.cnki.com.cn/Article/CDMD-10298-2005103040.htm [4] 李云峰, 曹渝昆, 朱庆生, 等.基于小波域隐马模型的树木类图像分割算法[J].计算机应用研究, 2007, 24(8): 233-235. doi: 10.3969/j.issn.1001-3695.2007.08.076Li Y F, Cao Y K, Zhu Q S, et al. Tree image segmentation algorithm based on wavelet-domain hidden-maze model[J]. Research of Computer Applications, 2007, 24(8):233-235. doi: 10.3969/j.issn.1001-3695.2007.08.076 [5] 张娜.树木图像分割方法的研究[D].哈尔滨: 东北林业大学, 2013. http://cdmd.cnki.com.cn/Article/CDMD-10225-1013371465.htmZhang N.Research on tree image segmentation method[D]. Harbin: Northeast Forestry University, 2013. http://cdmd.cnki.com.cn/Article/CDMD-10225-1013371465.htm [6] 程磊.树木图象的分割方法初探[D].北京: 北京林业大学, 2004. http://cdmd.cnki.com.cn/article/cdmd-10022-2004116979.htmCheng L. The method of tree image segmentation[D]. Beijing: Beijing Forestry University, 2004. http://cdmd.cnki.com.cn/article/cdmd-10022-2004116979.htm [7] 郭晶晶, 李庆武, 程海粟, 等.基于Lab颜色距离和GMM的树木图像分割算法[J].信息技术, 2016(2):1-4, 9. http://d.old.wanfangdata.com.cn/Periodical/xxjs201602001Guo J J, Li Q W, Cheng H S, et al. Tree image segmentation algorithm based on lab color distance and GMM[J]. Information Technology, 2016(2):1-4, 9. http://d.old.wanfangdata.com.cn/Periodical/xxjs201602001 [8] 王晓松, 黄心渊.一种改进的基于MRF的树木图像提取方法[J].北京林业大学学报, 2012, 34(5):128-133. http://j.bjfu.edu.cn/article/id/9811Wang X S, Huang X Y. An improved MRF based tree image extraction method[J].Journal of Beijing Forestry University, 2012, 34(5):128-133. http://j.bjfu.edu.cn/article/id/9811 [9] Ye N, Li X Y. A scalable, incremental learning algorithm for classification problems[J]. Computers & Industrial Engineering, 2002, 43(4): 677-692. http://cn.bing.com/academic/profile?id=42c70044a0b7abac36225de930766425&encoded=0&v=paper_preview&mkt=zh-cn [10] Musavi M T, Ahmed W, Chan K H, et al. On the training of radial basis function classifiers[J]. Neural Networks, 1992, 5(4): 595-603. doi: 10.1016/S0893-6080(05)80038-3 [11] Kim S W, Oommen B J. Enhancing prototype reduction schemes with LVQ3-type algorithms[J]. Pattern Recognition, 2003, 36(5): 1083-1093. doi: 10.1016/S0031-3203(02)00115-2 [12] Li X, Ye N. A supervised clustering and classification algorithm for mining data with mixed variables[J]. IEEE Transactions on Systems, Man, and Cybernetics, Part A: Systems and Humans, 2006, 36(2): 396-406. doi: 10.1109/TSMCA.2005.853501 [13] Setnes M, Babuska R. Fuzzy relational classifier trained by fuzzy clustering[J]. IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetic, 1999, 29(5): 619-625. doi: 10.1109/3477.790444 [14] Cai W L, Chen S C, Zhang D Q. A simultaneous learning framework for clustering and classification[J]. Pattern Recognition, 2009, 42(7): 1248-1259. doi: 10.1016/j.patcog.2008.11.029 [15] Cai W L, Chen S C, Zhang D Q. A multiobjective simultaneous learning framework for clustering and classification[J]. IEEE Transactions on Neural Networks, 2010, 21(2): 185-200. doi: 10.1109/TNN.2009.2034741 [16] Bezdek J C, Ehrlich R, Full W. FCM: the fuzzy c-means clustering algorithm[J]. Computers & Geosciences, 1984, 10(2/3): 191-203. http://d.old.wanfangdata.com.cn/Periodical/gpxygpfx201012036 [17] Coello C A C, Pulido G T, Lechuga M S. Handling multiple objectives with particle swarm optimization[J]. IEEE Transactions on Evolutionary Computation, 2004, 8(3): 256-279. doi: 10.1109/TEVC.2004.826067 [18] Deb K, Pratap A, Agarwal S, et al. A fast and elitist multiobjective genetic algorithm: NSGA-II[J]. IEEE Transactions on Evolutionary Computation, 2002, 6(2): 182-197. doi: 10.1109/4235.996017 [19] Wu C H, Ouyang C S, Chen L W, et al. A new fuzzy clustering validity index with a median factor for centroid-based clustering[J]. IEEE Transactions on Fuzzy Systems, 2015, 23(3): 701-718. doi: 10.1109/TFUZZ.2014.2322495 [20] Tang Y G, Sun F C, Sun Z Q. Improved validation index for fuzzy clustering[C]//Proceedings of the 2005, American Control Conference, 2005. Portland: IEEE, 2005, 2: 1120-1125. [21] Haouas F, Dhiaf Z B, Hammouda A, et al. A new efficient fuzzy cluster validity index: application to images clustering[C]//2017 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE). Naples: IEEE, 2017. [22] Deb K, Agrawal R B. Simulated binary crossover for continuous search space[J]. Complex Systems, 1994, 9(3): 115-148. http://cn.bing.com/academic/profile?id=fc75f2c786e997992685ce3da77fbd15&encoded=0&v=paper_preview&mkt=zh-cn [23] Deb K, Goyal M. A combined genetic adaptive search (GeneAS) for engineering design[J]. Computer Science Informatics, 1996, 26(4): 30-45. http://cn.bing.com/academic/profile?id=ca0c97135300a100909d6ad8af227e7b&encoded=0&v=paper_preview&mkt=zh-cn [24] Maulik U, Bandyopadhyay S. Performance evaluation of some clustering algorithms and validity indices[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2002, 24(12): 1650-1654. doi: 10.1109/TPAMI.2002.1114856 -
点击查看大图
图(4) / 表(2)
计量
- 文章访问数: 726
- HTML全文浏览量: 293
- PDF下载量: 10
- 被引次数: 0
