Page 244 - 应用声学2019年第4期
P. 244
704 2019 年 7 月
表 1 随机森林和 SVM 对三类目标的分类结果
(仿真样本) 参 考 文 献
Table 1 Random forest and SVM classi-
fication results for three kinds of targets [1] Kim S D, Lee J H, Kim J K. A new chain-coding algorithm
(simulated sample) for binary images using run-length codes[J]. Computer
Vision, Graphics, and Image Processing, 1988, 41(1):
随机森林 随机森林 114–128.
类别 SVM
(Gini) (Entropy) [2] Belkasim S O, Shridhar M, Ahmadi M. Pattern recog-
训练集的准确率 96.13% 95.91% 100.00% nition with moment invariants: a comparative study
测试集的准确率 95.50% 94.50% 94.00% and new results[J]. Pattern Recognition, 1991, 24(12):
1117–1138.
[3] Barr A H. Superquadrics and angle-preserving transfor-
表 2 随机森林和 SVM 对三类目标的分类结果
mations[J]. IEEE Computer Graphics and Applications,
(实际样本) 1981, 1(1): 11–23.
Table 2 Random forest and SVM classi- [4] Pentland A P. Perceptual organization and the repre-
fication results for three kinds of targets sentation of natural form[J]. Artificial Intelligence, 1986,
28(3): 293–331.
(actual sample)
[5] Peak P. Mathematical games—The superellipse: a curve
that lies between the ellipse and the rectangle, scien-
随机森林 随机森林
类别 SVM tific american by martin gardner[J]. Mathematics Teacher,
(Gini) (Entropy)
1966, 59(5): 473.
训练集的准确率 95.27% 94.14% 100.00%
[6] 陈京, 袁保宗, 文富荣. 一种基于曲率约束的不完整超二次曲
测试集的准确率 86.67% 78.57% 80.00% 线拟合 [C]//第十二届全国信号处理学术年会 (CCSP-2005)
论文集, 2005: 367–370.
从表 2可以看出,对于实际样本,三种分类器的 [7] Rosin P L. Fitting superellipses[J]. IEEE Transactions
分类结果均低于仿真样本,由于实际的水下声图像 on Pattern Analysis & Machine Intelligence, 2000, 22(7):
726–732.
比较复杂,易于受噪声、海底混响等因素的影响,使
[8] Ahn S J, Rauh W, Cho S H, et al. Orthogonal distance
提取的目标阴影边界坐标点与实际的坐标点之间 fitting of implicit curves and surfaces[J]. IEEE Transac-
存在偏差,导致分类准确度下降。尽管分类准确度 tions on Pattern Analysis & Machine Intelligence, 2002,
24(5): 620–638.
有所下降,但是最高也能达到 86.67%,具有较好的
[9] Voss K, Suesse H. A new one-parametric fitting method
分类效果。 for planar objects[J]. IEEE Transactions on Pattern Anal-
通过分析样本集的特征向量,得出产生错分的 ysis & Machine Intelligence, 1999, 21(7): 646–651.
[10] Rosin P L, West G A W. Curve segmentation and repre-
原因主要集中在球体目标和圆台目标之间,当成像
sentation by superellipses[J]. IEE Proceedings-Vision, Im-
设备探测的角度和方位正好处在球体垂直直径位 age and Signal Processing, 1995, 142(5): 280–288.
置上,产生的阴影区域可能与圆台目标的阴影区域 [11] Fitzgibbon A, Pilu M, Fisher R B. Direct least square fit-
ting of ellipses[J]. IEEE Transactions on Pattern Analysis
相似,导致超椭圆曲线的参数之间区分度不大,从而
and Machine Intelligence, 1999, 21(5): 476–480.
降低分类准确度。 [12] Nelder J A, Mead R. A simplex method for function min-
imization[J]. Compute Journal, 1965, 7(4): 308–331.
4 结论 [13] Mignotte M, Collet C, Pérez P, et al. Three-class Marko-
vian segmentation of high-resolution sonar images[J].
Computer Vision and Image Understanding, 1999, 76(3):
本文针对水下小目标分类问题,引入了超椭圆
191–204.
曲线拟合算法来识别目标阴影轮廓特性,并将控制 [14] Kim S D, Lee J H, Kim J K. A new chain-coding algorithm
曲线形状的 6 个参数 a、b、x、y、θ、ε 作为特征向量输 for binary images using run-length codes[J]. Computer
Vision, Graphics, and Image Processing, 1988, 41(1):
入到分类器内进行分类。结果表明,a、b、x、y、θ、ε 可
114–128.
以有效地分类出目标所属的类别,因此,超椭圆曲线 [15] Cutler A, Cutler D R, Stevens J R. Random forests[J].
拟合算法是一种可行的识别方法。但是,这种方法 Machine Learning, 2004, 45(1): 157–176.
[16] Zhao Q, Principe J C. Support vector machines for SAR
对于极不规则的目标阴影形状识别效果不尽人意,
automatic target recognition[J]. IEEE Transactions on
从而降低分类准确率,仍需要进一步研究与改进。 Aerospace & Electronic Systems, 2001, 37(2): 643–654.