-
摘要: 针对目前人为探察土地资源利用情况的任务繁重、办事效率低下等问题, 提出了一种基于深度卷积神经网络的建筑物变化检测方法, 利用高分辨率遥感图像实时检测每个区域新建与扩建的建筑物, 以方便对土地资源进行有效管理.本文受超列(Hypercolumn)和FlowNet中的细化(Refinement)结构启发, 将细化和其他改进应用到U-Net, 提出FlowS-Unet网络.首先对遥感图像裁剪、去噪、标注语义制作数据集, 将该数据集划分为训练集和测试集, 对训练集进行数据增强, 并根据训练集图像的均值和方差对所有图像进行归一化; 然后将训练集输入集成了多尺度交叉训练、多重损失计算、Adam优化的全卷积神经网络FlowS-Unet中进行训练; 最后对网络模型的预测结果进行膨胀、腐蚀以及孔洞填充等后处理得到最终的分割结果.本文以人工分割结果为参考标准进行对比测试, 用FlowS-Unet检测得到的F1分数高达0.943, 明显优于FCN和U-Net的预测结果.实验结果表明, FlowS-Unet能够实时准确地将新建与扩建的建筑物变化检测出来, 并且该模型也可扩展到其他类似的图像检测问题中.
-
关键词:
- FlowS-Unet /
- 建筑物变化检测 /
- 全卷积神经网络 /
- 多尺度交叉训练 /
- 多重损失
Abstract: Since manually detecting the situation of land resource utilization is arduous and inefficient, a smart building change detection method based on deep convolutional network is proposed, which can detect newly emerged or expanded buildings in each region of the high-resolution remote sensing images at real-time, thus can be used to manage the land resources efficiently. This article proposes a model named FlowS-Unet by applying refinement and other improvements to U-Net, which was inspired by hypercolumns and the refinement structure in FlowNet. First, the remote sensing images were cropped, denoised, and semantically annotated to form the dataset which is further divided into the training set and testing set, the training set is augmented to get enough training samples, and the mean value and variance of all training images are calculated and used to normalize the dataset; Second, the training set is fed into the fully convolutional network FlowS-Unet for training, which integrates multi-scale cross training, multiple losses and Adam algorithm for its optimization. Finally, the predicted result of FlowS-Unet is further post-processed with dilating, eroding and hole-filling to get the final segmentation result. By using manually segmented results as the ground truth, a comparison with several different algorithms shows that the F1 score of FlowS-Unet is as high as 0.943, which is apparently better than the predicted results of fully convolutional networks (FCN) and U-Net. Experimental results indicate that the newly emerged or expanded buildings can be accurately detected at real time with FlowS-Unet. This model can also be applied to other similar image detection problems.-
Key words:
- FlowS-Unet /
- change detection for buildings /
- fully convolutional networks (FCN) /
- multi-scale cross training /
- multiple losses
-
$ H_{\infty} $控制理论主要研究抑制干扰和不确定性问题[1].在$ H_{\infty} $控制理论中, 传递函数(或系统)的$ H_{\infty} $范数是一项重要的性能指标, 用于度量扰动输入对系统输出的影响, 反映了闭环系统的抗扰能力.在$ H_{\infty} $控制理论研究中, 长期存在一个挑战性议题:是否能够直接给出关于$ H_{\infty} $范数的通用解析表达式, 进而避免针对线性矩阵不等式(Linear matrix inequality, LMI)约束条件的繁琐的$ H_{\infty} $范数近似寻优方案.
在20世纪80年代, $ H_{\infty} $控制理论的研究由频域转换到时域, 开启了基于状态空间方程描述的系统鲁棒性能研究[2].总的来说, $ H_{\infty} $性能时域分析面临的核心问题是如何选择适当的李雅普诺夫函数.具体表现为基于李雅普诺夫方程[3-4]或参数化Riccati不等式[5]均难以得到用于精确分析系统$ H_{\infty} $性能的最优李雅普诺夫函数, 因此在早期的研究中结果的保守性是难以避免的.
为精确求解$ H_{\infty} $范数, 有学者提出了有界实引理[6], 并将求解$ H_{\infty} $范数问题转化为时域状态空间的约束优化问题.基于有界实引理给出的LMI约束条件, $ H_{\infty} $范数能够被近似寻优[7-14].在LMI方法中, $ H_{\infty} $范数的寻优一般包含以下步骤:
1) 给出一个充分大的初始$ H_{\infty} $范数估计$ \mit\gamma $;
2) 解LMI问题;
3) 递减$ H_{\infty} $范数估计$ \mit\gamma $, 直到获得满足LMI条件的最小$ H_{\infty} $范数估计$ \mit\gamma $.
显然, 一旦最小$ H_{\infty} $范数估计得到, 则通过解LMI, 可以得到相应的近似最优李雅普诺夫函数.不难发现, LMI方法存在一定不足, 表现为:
1) 对于每一个给定的$ \mit\gamma $, LMI条件需要被重复求解, 直到找到最小的$ H_{\infty} $范数估计, 过程过于繁琐;
2) 这种试凑逼近方法无法揭示系统结构和参数对$ H_{\infty} $性能的影响, 在一定程度上限制了控制器精细设计的研究.
为了克服目前关于$ H_{\infty} $范数问题研究的不足, 一个可替换的方法是直接优化李雅普诺夫函数, 进而得到关于$ H_{\infty} $范数的通用解析表达式.目前, 针对系统具体性能, 难以找到李雅普诺夫函数设计的充要条件, 因此这方面的研究并不多见.事实上, 在分析系统具体性能时, 存在最优的李雅普诺夫函数, 并且这一最优李雅普诺夫函数与系统结构和参数存在内在关系[15].因此本文尝试寻找一种李雅普诺夫函数的直接优化途径, 进而实现$ H_{\infty} $性能的精确分析.
由于多数高阶系统在一定的条件下可以近似(或分解)为二阶系统来研究, 并且二阶系统的分析方法是分析高阶系统的基础[16], 因此为有效展现最优李雅普诺夫函数与系统结构和参数存在内在关系, 本文针对一类二阶系统的$ H_{\infty} $范数问题, 构造和优化李雅普诺夫函数, 进而得到$ H_{\infty} $范数的通用解析表达式.本文的研究避免了LMI方法中繁琐的近似寻优过程, 并展示了系统矩阵特征值的实部和虚部对$ H_{\infty} $性能的影响.本文结构如下:第1节分析$ H_{\infty} $范数问题; 第2节分析Riccati不等式中李雅普诺夫函数的选择对求解$ H_{\infty} $范数的影响; 第3节展现李雅普诺夫函数的直接优化方法, 并给出$ H_{\infty} $范数的通用解析表达式; 第4节给出算例, 验证李雅普诺夫函数直接优化方法的有效性.
1. 问题的提出
1.1 问题描述
系统描述为
$ \begin{align} \dot{\boldsymbol{ x}} = A {\boldsymbol{ x}}+ {\boldsymbol{ w}} \end{align} $
(1) 其中, $ {\boldsymbol{ x}} \in \textbf{R}^{2} $, $ A $为Hurwitz矩阵, $ A $的特征值为复数, $ {\boldsymbol{ w}} $为扰动输入, $ \|{\boldsymbol{ w}}\| \leq \delta $, $ \delta $为常数, $ \|{\boldsymbol{ w}}\| = (\Sigma^{2}_{i = 1}w^{2}_{i})^{\frac{1}{2}} $.
研究的问题是如何得到系统(1)的状态上界.在数学意义上, 这一问题可转化为关于输入–输出系统的$ H_{\infty} $范数问题, 其中系统描述为
$ \begin{align} \begin{cases} \dot{\boldsymbol{ x}} = A {\boldsymbol{ x}} + {\boldsymbol{ w}} \\ {\boldsymbol{ y}} = {\boldsymbol{ x}} \end{cases} \end{align} $
(2) 在$ H_{\infty} $控制理论中, 系统的$ H_{\infty} $范数定义为$ S $右半平面上解析的有理函数阵的最大奇异值.在标量函数中就是幅频特性的极大值, 代表了系统对峰值有界信号的传递特性.
1.2 LMI方法分析
令李雅普诺夫函数为$ V = {\boldsymbol{ x}}^{\rm T}P{\boldsymbol{ x}} $, $ \gamma $为系统(2)的$ H_{\infty} $范数, 即$ \mit\gamma = \|G\|_{\infty} $, 其中$ G(s) = (sI-A)^{-1} $为系统(2)的传递函数.根据有界实引理, 可得
$ \begin{align} \left[ \begin{array}{ccc} PA+A^{\rm{T}}P & P & I \\ P & -\gamma^{2} I & 0_{2\times 2} \\ I & 0_{2\times 2} & -I \\ \end{array} \right] < 0 \end{align} $
(3) LMI方法是寻找式(3)中$ \mit\gamma $的最小值$ \mit\gamma_{\rm{min}} $.由于李雅普诺夫函数$ V = {\boldsymbol{ x}}^{\rm T}P {\boldsymbol{ x}} $可以任意构造, 因此对于每一个给定的$ \mit\gamma $, 需要重复求解LMI, 以判断式(3)的存在性, 直到$ \mit\gamma_{\rm{min}} $被找到.显然, 在LMI方法中复杂的优化过程是不可避免的.事实上, $ \mit\gamma_{\rm{min}} $与最优的$ P $矩阵是一一对应的.如果能够直接给出最优的$ P $矩阵, 则$ \mit\gamma_{\rm{min}} $的表达式就能够得到, 进而避免LMI方法中复杂的优化过程.本文的工作是尝试提供一种新的途径来直接给出$ \mit\gamma_{\rm{min}} $的表达式.
2. $ \pmb H_{\boldsymbol{ \infty}} $范数分析
根据特征值和奇异值分解原理, 可以得到下面的特性.
特性1. 对于系统(2)中特征矩阵$ A $, 存在可逆矩阵$ T $, 满足
$ \begin{align} D = -TAT^{-1} = \left[ \begin{array}{cc} \lambda & \nu \\ -\nu & \lambda \\ \end{array} \right] \end{align} $
(4) 其中, $ T = \Theta_{T1} \times \text{diag}\{t_{1}, t_{2}\} \times \Theta_{T2} $, $ \Theta_{T1} $和$ \Theta_{T2} $为正交矩阵, $ t_{2} \geq t_{1} > 0 $, $ \lambda > 0 $, $ \nu > 0 $. $ \text{diag}\{t_{1}, t_{2}\} $表示对角元素为$ t_{1} $, $ t_{2} $的对角阵.
令$ \alpha = {t_{2}}/{t_{1}} \geq 1 $, $ {\boldsymbol{ y}} = \Theta_{T2} \times {\boldsymbol{ x}} $, $ {\boldsymbol{ {\Delta}}} = \Theta_{T2}\times{\boldsymbol{ w}} $.由式(2)和特性1, 得
$ \begin{align} \begin{cases} \dot{\boldsymbol{ y}} = E {\boldsymbol{ y}} + B {\boldsymbol{ {\Delta}}} \\ {\boldsymbol{ x}} = C {\boldsymbol{ y}} \end{cases} \end{align} $
(5) 其中, $ B = I $为单位阵, $ C = \Theta_{T2}^{-1} $, $ E = - \left[ {array}{cc} \lambda & \alpha \nu \\ -\frac{1}{\alpha}\nu & \lambda \\ {array} \right], $并且系统(2)和(5)具有相同的$ H_{\infty} $范数.
根据文献[5]中引理2.1, 可以得到下面的特性.
特性2. 对于系统(5), 存在正定矩阵$ X $, 满足Riccati不等式
$ \begin{align} E^{\rm T}X+XE+(1+\varepsilon)C^{\rm T}C+ \rho^{-2} XBB^{\rm T}X \leq 0 \end{align} $
(6) 其中, $ \gamma < \rho $, $ \gamma = \|G\|_{\infty} $为系统$ H_{\infty} $范数, $ \varepsilon $为趋于零的正数.
注1. 应用Riccati不等式一般会得到具有很强保守性的结果, 但这种保守性并不是Riccati不等式本身导致的.研究表明:基于李雅普诺夫函数的准确选择, 可以将特性2中Riccati不等式转化为等式, 进而精确给出$ H_{\infty} $范数.因此, 导致这种保守性的原因是:在应用Riccati不等式时, 目前尚没有有效的方法找到最优的李雅普诺夫函数.这正是本文研究李雅普诺夫函数构造(或优化)的动机.
令
$ \begin{align} \Upsilon = \, &K^{-1} \Theta \begin{bmatrix} \lambda & -\frac{1}{\alpha} \nu \\ \alpha \nu & \lambda \end{bmatrix}\Theta^{\rm T}\; + \nonumber \\&\Theta \begin{bmatrix} \lambda & \alpha \nu \\ -\frac{1}{\alpha} \nu & \lambda \\ \end{bmatrix} \Theta^{\rm T}K^{-1} - K^{-1}K^{-1} \end{align} $
(7) 其中, $ \alpha \geq 1 $,
$ \begin{align} K = \iota \left[ \begin{array}{cc} 1 & 0 \\ 0 & k \\ \end{array} \right], \;\;\;\; \Theta = \left[ \begin{array}{cc} \cos \theta & \sin \theta \\ -\sin \theta & \cos \theta \\ \end{array} \right] \end{align} $
(8) $ \iota >0 $, $ k \geq 1 $, $ 0 \leq \theta \leq {\pi}/{4} $.
由式(8)构造的李雅普诺夫函数分解了"放缩"和"旋转"作用.这种功能的分解使李雅普诺夫函数的参数优化具有了可行性.
定理1. 对于系统(5), 系统$ H_{\infty} $范数$ \gamma $满足
$ \begin{align} \gamma < \rho_{\rm{min}} = \left[\sqrt{\lambda_{\rm{min}}(\Upsilon)} \right]^{-1} \end{align} $
(9) 其中, $ \lambda_{\rm{min}}(\Upsilon) $为矩阵$ \Upsilon $的最小特征值.
证明. 令$ X = \Theta^{\rm T} K \Theta $, 其中, $ K $和$ \Theta $由式(8)给出.根据特性2和式(7), 得
$ \begin{align} \rho^{-2} I \leq \Upsilon - \varepsilon K^{-1}K^{-1} \end{align} $
(10) 则$ \rho^{-2} \leq \lambda_{\rm{min}}(\Upsilon- \varepsilon K^{-1}K^{-1}) $, 由于$ \gamma < \rho $, 并且$ \varepsilon $为趋于零的正数, 则式(9)成立.
注2. 根据定理1, 可以优化李雅普诺夫函数的参数, 以最大化$ \lambda_{\rm{min}}(\Upsilon) $, 进而精确估计系统$ H_{\infty} $范数.因此, 定理1给出了一种新的途径以得到系统的$ H_{\infty} $范数.
3. 李雅普诺夫函数优化
考查式(7)给出的矩阵$ \Upsilon $.由式(7)和式(8), 可得
$ \begin{align} \Upsilon = \frac{1}{\iota} \left[ \begin{array}{cc} 2\lambda + \beta \nu - \frac{1}{\iota} & \frac{1}{k} \sigma \nu \\ \frac{1}{k} \sigma \nu & \frac{1}{k}(2 \lambda - \beta \nu) - \frac{1}{\iota k^{2}} \\ \end{array} \right] \end{align} $
(11) 其中,
$ \begin{align} \beta = &\ \left(\alpha-\frac{1}{\alpha}\right) \sin 2\theta \end{align} $
(12) $ \begin{align} \sigma = &\, \left[\alpha- (\alpha-\frac{1}{\alpha}) \sin^{2} \theta \right] -k \left[\frac{1}{\alpha} + (\alpha-\frac{1}{\alpha}) \sin^{2} \theta \right] = \\ &\ \frac{1}{2}(1-k)(\alpha+\frac{1}{\alpha}) +\frac{1}{2}(1+k) (\alpha-\frac{1}{\alpha}) \cos 2\theta \end{align} $
(13) 根据式(11), 以最大化$ \lambda_{\rm{min}}(\Upsilon) $为目标, 将给出一种李雅普诺夫函数的优化方法.
3.1 李雅普诺夫函数优化策略
令
$ \begin{align} \Upsilon_{1} = \Theta^{-1} \Upsilon \Theta, \; \; Y_{1} = X^{-1} \end{align} $
(14) 则由式(7)和$ X = \Theta^{\rm T}K\Theta $, 得
$ \begin{align} \Upsilon_{1} = EE^{\rm T}-(E+Y_{1})(E+Y_{1})^{\rm T} \end{align} $
(15) 令
$ \begin{align} &EE^{\rm T} = \Theta_{1}^{\rm T} \Lambda \Theta_{1}, \quad \Upsilon_{2} = \Theta_{1} \Upsilon_{1} \Theta_{1}^{\rm T} \end{align} $
(16) $ \begin{align} &E_{1} = \Theta_{1} E \Theta_{1}^{\rm T}, \qquad Y_{2} = \Theta_{1} Y_{1} \Theta_{1}^{\rm T} \end{align} $
(17) 其中, $ \Lambda = {\rm diag}\{\sigma_{1}, \sigma_{2}\} $, $ \sigma_{1} \geq \sigma_{2} $, 则
$ \begin{align} \Upsilon_{2} = \Lambda - (E_{1}+Y_{2})(E_{1}+Y_{2})^{\rm T} \end{align} $
(18) 令
$ \begin{align} E_{1} = E_{R}+E_{J}, \; \; Y_{3} = E_{R}+Y_{2} \end{align} $
(19) 其中, $ E_{R}^{\rm T} = E_{R} $, $ E_{J} = -E_{J}^{\rm T} $, 则
$ \begin{align} \Upsilon_{2} = \Lambda - (E_{J}+Y_{3})(E_{J}+Y_{3})^{\rm T} \end{align} $
(20) 令
$ \begin{align} Y_{3} = \left[ \begin{array}{cc} y_{1} & y_{3} \\ y_{3} & y_{2} \\ \end{array} \right], \; \; E_{J} = \left[ \begin{array}{cc} 0 & a \\ -a & 0 \\ \end{array} \right] \end{align} $
(21) 则根据$ \Lambda = \text{diag}\{\sigma_{1}, \sigma_{2}\} $, 有$ \sigma_{1} \geq \sigma_{2} $,
$ \begin{align} \Upsilon_{2} = & \left[ \begin{array}{cc} \sigma_{1}-(y_{3}+a)^{2}-y_{1}^{2} \\ -(y_{1}+y_{2})y_{3}-(y_{2}-y_{1})a \\ \end{array}\right.\\ &\qquad\qquad\qquad \left. \begin{array}{cc} & -(y_{1}+y_{2})y_{3}-(y_{2}-y_{1})a \\ & \sigma_{2} -(y_{3}-a)^{2}-y_{2}^{2} \\ \end{array} \right] \end{align} $
(22) 根据式(14), (16), (21), (22)和定理1, 存在$ Y_{3} $, 使$ \lambda_{\rm{min}}(\Upsilon_{2}) $ $ > $ $ 0 $, 即$ \Upsilon_{2} $正定.因此根据式(22), 为了最大化$ \Upsilon_{2} $的最小特征值, 应使下面两个条件成立.
1) $ (y_{1}+y_{2})y_{3}+ (y_{2}-y_{1})a = 0 $ (例如$ y_{2} = 0 $, $ y_{3} = a $; 或$ y_{1} = y_{2} = 0 $).
2) $ \Upsilon_{2} $的特征值相等(例如$ y_{1}^{2} = \sigma_{1}-\sigma_{2}-4a^{2} $; 或$ y_{3} $ $ = $ $ (\sigma_{1}-\sigma_{2})/{4a} $).
注意, $ \sqrt{\sigma_{2}} $为$ E $的最小奇异值, 因此$ \gamma \geq {1}/{\sqrt{\sigma_{2}}} $.令
$ \begin{align} \lambda_{1} = \frac{1}{\iota}\left( 2\lambda + \beta \nu - \frac{1}{\iota} \right), \; \; \lambda_{2} = \frac{1}{\iota}\left[ \frac{1}{k}(2 \lambda - \beta \nu) - \frac{1}{\iota k^{2}} \right] \end{align} $
(23) 基于以上分析, 并根据式(9), (11), (14), (16)和(23), 为了最大化$ \Upsilon $的最小特征值, 李雅普诺夫函数的优化策略设计为$ \sigma = 0 $和$ \lambda_{1} = \lambda_{2} $.
3.2 李雅普诺夫函数参数优化
基于所给李雅普诺夫函数优化策略, 进一步优化李雅普诺夫函数参数.
定理2. 对于系统(5), 系统$ H_{\infty} $范数$ \gamma $满足
$ \begin{align} \gamma < \rho(k, \iota) = \left[\min(\lambda_{1}, \lambda_{2}) \right]^{-\frac{1}{2}} \end{align} $
(24) 其中, $ \lambda_{1} $和$ \lambda_{2} $由式(23)给出, 式(23)中$ \beta $由下式给出.
$ \begin{align} \beta = \frac{2}{k+1}\sqrt{\left(k \alpha-\frac{1}{\alpha}\right)\left(\alpha- \frac{k}{\alpha}\right)} \end{align} $
(25) 证明. 考查式(11)给出的矩阵$ \Upsilon $.令$ \sigma = 0 $, 则
$ \begin{align} \cos 2\theta = \frac{(k-1)(\alpha+\frac{1}{\alpha})}{(k+1)(\alpha-\frac{1}{\alpha})} \end{align} $
(26) 因此根据式(11), (12), (23)和$ 0 \leq \theta \leq {\pi}/{4} $, 矩阵$ \Upsilon $的特征值为$ \lambda_{1} $和$ \lambda_{2} $, 其中$ \beta $由式(25)给出.根据定理1, 可得式(24).
注3. 基于李雅普诺夫函数参数矩阵$ \Theta $的优化策略, 定理2进一步给出系统$ H_{\infty} $范数的估计., 同时奠定了进一步优化李雅普诺夫函数参数$ k $和$ \iota $的基础.
定理3. 对于系统(5), 系统$ H_{\infty} $范数$ \gamma $满足
$ \begin{align} \gamma < \rho(k) = \begin{cases} \frac{1}{\lambda}, & \text{若}\; \alpha = 1\\ \left[ f(k)\right]^{-\frac{1}{2}}, & \text{若}\; \alpha >1 \end{cases} \end{align} $
(27) 其中,
$ \begin{align} f(k) = \frac{4k}{(k+1)^{2}} \left[ \lambda^{2} + \nu^{2} - \frac{k \nu^{2}}{(k-1)^{2}} \left(\alpha-\frac{1}{\alpha}\right)^{2} \right] \end{align} $
(28) 证明. 考查式(23)给出的矩阵$ \Upsilon $的特征值为$ \lambda_{1} $和$ \lambda_{2} $.令$ \lambda_{1} = \lambda_{2} $, 即
$ \begin{align} 2\lambda + \beta \nu - \frac{1}{\iota} = \frac{1}{k}(2 \lambda - \beta \nu) - \frac{1}{\iota k^{2}} \end{align} $
(29) 其中, $ \beta $由式(25)给出, $ \alpha \geq 1 $.
当$ \alpha > 1 $时, 由式(25)和式(29)可知$ k \neq 1 $, 并且得
$ \begin{align} \frac{1}{\iota} = \frac{2k \lambda}{k+1}+\frac{2k \nu}{k^{2}-1} \sqrt{\left(k \alpha- \frac{1}{\alpha}\right)\left(\alpha-\frac{k}{\alpha}\right)} \end{align} $
(30) 当$ \alpha = 1 $时, 由式(25)可知$ (k-1)^{2} \leq 0 $, 即$ k = 1 $.则根据式(23), (25), (29), $ \lambda_{1} = \lambda_{2} = \frac{1}{\iota} (2 \lambda-\frac{1}{\iota}) $.当$ \iota = \lambda $时, 得$ \max (\lambda_{1}) = \lambda^{2} $.
基于以上分析, 并根据定理2和式(23), (25), (29)以及(30), 可得结论.
注4. 通过给出李雅普诺夫函数参数$ \iota $的优化策略, 定理3进一步给出系统$ H_{\infty} $范数的估计.根据定理3, 可以直接优化李雅普诺夫函数参数$ k $, 进而得到系统$ H_{\infty} $范数的精确估计.
注5. 注意, 当$ \alpha > 1 $时, $ k \neq 1 $.因此定理3通过分别讨论$ \alpha > 1 $和$ \alpha = 1 $两种情况, 解决了$ f(k) $的奇异问题.
令
$ \begin{align} \kappa = k + \frac{1}{k} > 2 \end{align} $
(31) 则由式(28), 得
$ \begin{align} f(\kappa) = \frac{4(\lambda^{2} + \nu^{2})}{\kappa+2} - \frac{4\nu^{2}}{\kappa^{2}-4} \times \left(\alpha-\frac{1}{\alpha}\right)^{2} \end{align} $
(32) 定理4. 对于系统(5), 系统$ H_{\infty} $范数$ \gamma $满足
$ \begin{align} \gamma < \rho_{\text{opt}} = \begin{cases} \frac{1}{\lambda}, & \text{若}\; \alpha = 1\\ \frac{1}{2\lambda}\sqrt{\alpha^{2}+\frac{1}{\alpha^{2}}+2}, &\text{若}\; \kappa_{0} \geq \alpha^{2}+\frac{1}{\alpha^{2}}\\ \left[ f(\kappa_{0})\right]^{-\frac{1}{2}}, &\text{若}\; \kappa_{0} < \alpha^{2}+\frac{1}{\alpha^{2}} \end{cases} \end{align} $
(33) 其中
$ \begin{align} &f(\kappa_{0}) = \frac{4(\lambda^{2} + \nu^{2})}{\kappa_{0}+2} - \frac{4\nu^{2}}{\kappa_{0}^{2}-4} \times \left(\alpha-\frac{1}{\alpha}\right)^{2} \end{align} $
(34) $ \begin{align} &\kappa_{0} = 2 + \frac{\nu^{2} (\alpha-\frac{1}{\alpha})^{2}}{\lambda^{2} + \nu^{2}} \times \left[ 1+\sqrt{1+ \frac{4(\lambda^{2} + \nu^{2})}{\nu^{2} (\alpha-\frac{1}{\alpha})^{2}}} \right] \end{align} $
(35) 证明. 由式(32), 得
$ \begin{align} f'(\kappa) = \frac{{\rm d} f(\kappa)}{{\rm d} \kappa} = -\frac{4(\lambda^{2} + \nu^{2})}{(\kappa+2)^{2}} +\frac{8(\alpha-\frac{1}{\alpha})^{2} \nu^{2} \kappa}{(\kappa+2)^{2}(\kappa-2)^{2}} \end{align} $
(36) 令$ f'(\kappa) = 0 $, 即
$ \begin{align} \kappa^{2} - \left[ 4+ \frac{2(\alpha-\frac{1}{\alpha})^{2} \nu^{2}}{\lambda^{2} + \nu^{2}} \right] \kappa +4 = 0 \end{align} $
(37) 根据$ \kappa >2 $和式(35), 得$ \kappa = \kappa_{0} $.
根据式(35) $ \sim $ (37), 得
$ \begin{align} \lim \limits_{\varsigma \rightarrow 0} \frac{f'(\kappa_{0} + \varsigma)-f'(\kappa_{0})}{\varsigma} <0 \end{align} $
(38) 因此, 在$ 2 < \kappa < \infty $的条件下, $ \max f(\kappa) = f(\kappa_{0}) $, 如图 1 (a)和1 (b)所示.
注意, 定理2中李雅普诺夫函数参数矩阵$ \Theta $的优化策略为$ \sigma = 0 $, 则由式(13), 可得$ k \leq \alpha^{2} $.由于$ k >1 $, 因此根据式(31), 得
$ \begin{align} \Omega = \left\{ \kappa \in \textbf{R} | 2 < \kappa \leq \alpha^{2}+\frac{1}{\alpha^{2}} \right\} \end{align} $
(39) $ \begin{align} \max \limits_{k \in \Omega} f(\kappa) = \begin{cases} \frac{4\lambda^{2}}{\alpha^{2}+\frac{1}{\alpha^{2}}+2}, &\text{若}\; \kappa_{0} \geq \alpha^{2}+\frac{1}{\alpha^{2}}\\ f(\kappa_{0}), & \text{若}\; \kappa_{0} < \alpha^{2}+\frac{1}{\alpha^{2}} \end{cases} \end{align} $
(40) 因此由定理3可得结论.
注6. 通过对李雅普诺夫函数参数的直接优化, 定理4给出了系统$ H_{\infty} $范数上界的优化结果.应用定理4, 可以给出系统$ H_{\infty} $范数的精确估计.
注7. 不同于LMI方法, 本文提出的李雅普诺夫函数直接优化方法分析了李雅普诺夫函数的构造对系统性能分析的影响, 充分利用系统结构和参数以优化李雅普诺夫函数的设计.与LMI方法相比, 李雅普诺夫函数直接优化方法能够直接给出系统$ H_{\infty} $范数的精确结果, 进而避免了复杂的数值优化过程.因此本文的工作提供了一种新的途径以更为方便地分析系统动态性能.
4. 算例
考查系统
$ \begin{align} \dot{\boldsymbol{ x}} = -\left[ \begin{array}{cc} 1.25 & 1.25 \\ -1.25 & 2.75 \\ \end{array} \right]{\boldsymbol{ x}}+ {\boldsymbol{ w}} \end{align} $
(41) 其中, $ {\boldsymbol{ w}} $为扰动输入, $ \|{\boldsymbol{ w}}\| \leq 1 $, $ {\boldsymbol{ x}} $为状态输出.根据式(5), 得
$ \begin{align} \begin{cases} \dot{\boldsymbol{ y}} = - \left[ \begin{array}{cc} 2 & 2 \\ -0.5 & 2 \\ \end{array} \right] {\boldsymbol{ y}} + {\boldsymbol{ {\Delta}}} \\ {\boldsymbol{ x}} = \frac{\sqrt{2}}{2} \left[ \begin{array}{cc} 1 & -1 \\ 1 & 1 \\ \end{array} \right] {\boldsymbol{ y}} \end{cases} \end{align} $
(42) 因此, $ \lambda = 2 $, $ \nu = 1 $, $ \alpha = 2 $.
由式(34), 得$ \kappa_{0} = 3.8651< \alpha^{2}+\frac{1}{\alpha^{2}} = 4.25 $.则根据定理4, 得$ \gamma < \rho_{\text{opt}} = 0.622 $.因此$ \gamma \approx 0.622 $.应用MATLAB中$ H_{\infty} $范数求解函数hinfnorm (sys, 0.0000001)可得相同的结果.因此提出的李雅普诺夫函数直接优化方法能精确给出系统$ H_{\infty} $范数.
表 1进一步给出在不同参数条件下系统(5)的$ H_{\infty} $范数.表 1表明, 针对式(5)给出的具有不同参数的系统, 提出的李雅普诺夫函数直接优化方法都能精确给出系统$ H_{\infty} $范数.
表 1 $H_{\infty}$范数分析($\alpha = 2$)Table 1 $H_{\infty}$ norm analysis ($\alpha = 2$)$\lambda$ $\nu$ MATLAB 定理4 稳态误差$\|A^{-1}\|$ 状态上界 2 6 0.626 0.626 0.307 0.626 2 4 0.626 0.626 0.419 0.626 2 2 0.626 0.626 0.588 0.626 2 1.2 0.626 0.626 0.626 0.626 2 1 0.622 0.622 0.622 0.622 2 0 0.501 0.501 0.501 0.501 在$ \alpha $和系统特征值实部$ \lambda $确定(即$ \alpha = 2 $, $ \lambda = 2 $)的条件下, 表 1给出的结果表明, 随着系统特征值虚部$ \nu $变化, $ H_{\infty} $范数的变化具有一定规律性, 表现为:
1) 当$ \nu = \nu^{*} = 1.2 $ (即$ \kappa_{0} = \alpha^{2}+{1}/{\alpha^{2}} $)时, $ H_{\infty} $范数为$ \max \|A^{-1}\| $;
2) 当$ \nu < \nu^{*} $ (即$ \kappa_{0} < \alpha^{2}+{1}/{\alpha^{2}} $)时, $ H_{\infty} $范数与稳态指标$ \|A^{-1}\| $一致;
3) 当$ \nu > \nu^{*} $ (即$ \kappa_{0} > \alpha^{2}+{1}/{\alpha^{2}} $)时, $ H_{\infty} $范数为固定值(即$ H_{\infty} $范数的值与$ \nu $无关), 并且根据定理4, $ H_{\infty} $范数的表达式非常简洁.
由式(1), (3), (41), 得
$ \begin{align} \begin{bmatrix} -P \begin{bmatrix} 1.25 & 1.25 \\ -1.25 & 2.75 \\ \end{bmatrix} -\small{ \begin{bmatrix} 1.25 & -1.25 \\ 1.25 & 2.75 \\ \end{bmatrix}}P & P & I \\ P & -\gamma^{2} I & 0_{2\times 2} \\ I & 0_{2\times 2} & -I \end{bmatrix} < 0 \end{align} $
(43) 采用LMI方法求解$ H_{\infty} $范数的步骤为:
1) 选择足够大的$ \gamma $, 如$ \gamma = 10 $;
2) 应用MATLAB中LMI工具求解式(43), 可得$ P $存在;
3) 减小$ \gamma $取值, 如$ \gamma = 1 $, 应用LMI工具求解式(43), 可得$ P $存在;
4) 当$ \gamma = 0.622 $时, 应用LMI工具求解式(43), 可得$ P $存在;
5) 当$ \gamma = 0.621 $时, 应用LMI工具求解(43), 可得$ P $不存在.
基于以上步骤, LMI方法可给出$ H_{\infty} = 0.622 $.这一结果与定理4得到的结果一致, 如表 1所示.
事实上, LMI方法需要对$ \gamma $进行遍历寻找.当选$ \gamma $的间隔较大时, 保守的结果不可避免.与之相比, 本文的方法具有明显的优越性.
5. 结论
本文针对$ H_{\infty} $控制理论研究中难以精确求解系统$ H_{\infty} $范数的问题, 提出了一种李雅普诺夫函数的直接优化方法.通过优化Riccati不等式中的李雅普诺夫函数, 给出了$ H_{\infty} $范数的通用解析表达式, 进而提供了一个有效的途径以直接和精确求解系统$ H_{\infty} $范数.研究结果具有以下特点:
1) 与LMI方法相比, 本文所提方法避免了复杂的数值优化过程, 使求解系统$ H_{\infty} $范数简化.
2) 与早期关于李雅普诺夫方程和Riccati不等式的研究相比, 本文所提方法避免了由于李雅普诺夫函数选择的随意性导致的保守结果.
3) 本文所提方法能够展现系统矩阵特征值的实部和虚部对$ H_{\infty} $性能的影响, 为进一步精确(定量)控制系统$ H_{\infty} $性能提供借鉴.
在进一步的工作中, 将研究含有时滞及非线性项的系统.
-
表 1 FlowS-Unet与现有方法的性能比较
Table 1 The performance comparison of FlowS-Unet
序号 方法 F1分数(后处理前/后) 时间(s) 1 FlowS-Unet 0.933/0.943 62 2 FCN 0.858/0.873 50 3 U-Net 0.898/0.913 59 4 人工标注 1.000 18 000 表 2 多尺度与单尺度训练及预测的F1分数比较
Table 2 The F1 score comparison between multi-scale cross and single-scale training and testing
训练尺度(像素) 预测尺度(像素) F1分数(后处理前/后) 224 224 0.903/0.923 256 256 0.909/0.928 288 288 0.913/0.931 320 320 0.911/0.932 224 0.933/0.939 256 0.938/0.943 多尺度 288 0.939/0.945 320 0.939/0.944 多尺度平均 0.942/0.946 表 3 FlowS-Unet与其他队伍的F1分数比较
Table 3 The F1 score comparison of FlowS-Unet and other teams
名次 初赛 复赛 决赛 第1名 0.890 0.914 0.861 第2名(FlowS-Unet) 0.903 0.877 0.840 第3名 0.867 0.898 0.800 第4名 0.706 0.870 0.842 第5名 0.879 0.936 0.823 -
[1] Beumier C, Idrissa M. Building change detection from uniform regions. In: Proceedings of the 17th Iberoamerican Congress Pattern Recognition, Image Analysis, Computer Vision, and Applications. Buenos Aires, Argentina: Springer, 2012. 648-655 [2] Turker M, Sumer E. Building-based damage detection due to earthquake using the watershed segmentation of the post-event aerial images. International Journal of Remote Sensing, 2008, 29(11): 3073-3089 doi: 10.1080/01431160701442096 [3] Huang X, Zhang L P, Zhu T T. Building change detection from multitemporal high-resolution remotely sensed images based on a morphological building index. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2014, 7(1): 105-115 doi: 10.1109/JSTARS.2013.2252423 [4] 周则明, 孟勇, 黄思训, 胡宝鹏.基于能量最小化的星载SAR图像建筑物分割方法.自动化学报, 2016, 42(2): 279-289 doi: 10.16383/j.aas.2016.c150460Zhou Ze-Ming, Meng Yong, Huang Si-Xun, Hu Bao-Peng. Building segmentation of spaceborne SAR images based on energy minimization. Acta Automatica Sinica, 2016, 42(2): 279-289 doi: 10.16383/j.aas.2016.c150460 [5] 李炜明, 吴毅红, 胡占义.视角和光照显著变化时的变化检测方法研究.自动化学报, 2009, 35(5): 449-461 doi: 10.3724/SP.J.1004.2009.00449Li Wei-Ming, Wu Yi-Hong, Hu Zhan-Yi. Urban change detection under large view and illumination variations. Acta Automatica Sinica, 2009, 35(5): 449-461 doi: 10.3724/SP.J.1004.2009.00449 [6] 田昊, 杨剑, 汪彦明, 李国辉.基于先验形状约束水平集模型的建筑物提取方法.自动化学报, 2010, 36(11): 1502-1511 doi: 10.3724/SP.J.1004.2010.01502Tian Hao, Yang Jian, Wang Yan-Ming, Li Guo-Hui. Towards automatic building extraction: Variational level set model using prior shape knowledge. Acta Automatica Sinica, 2010, 36(11): 1502-1511 doi: 10.3724/SP.J.1004.2010.01502 [7] Liu B, Tang K, Liang J. A bottom-up/top-down hybrid algorithm for model-based building detection in single very high resolution SAR image. IEEE Geoscience and Remote Sensing Letters, 2017, 14(6): 926-930 doi: 10.1109/LGRS.2017.2687946 [8] Lukashevich P, Zalessky B, Belotserkovsky A. Building detection on aerial and space images. In: Proceedings of the 2017 International Conference on Information and Digital Technologies (IDT). Zilina, Slovakia: IEEE, 2017. 246-251 [9] 施文灶, 毛政元.基于图分割的高分辨率遥感影像建筑物变化检测研究.地球信息科学学报, 2016, 18(3): 423-432 http://d.old.wanfangdata.com.cn/Periodical/dqxxkx201603017Shi Wen-Zao, Mao Zheng-Yuan. The research on building change detection from high resolution remotely sensed imagery based on graph-cut segmentation. Journal of Geo-Information Science, 2016, 18(3): 423-432 http://d.old.wanfangdata.com.cn/Periodical/dqxxkx201603017 [10] Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks. In: Proceedings of the 25th International Conference on Neural Information Processing Systems. Nevada, USA: Curran Associates Inc., 2012. 1097-1105 [11] Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. In: Proceedings of the 3rd International Conference on Learning Representations. San Diego, California, USA, 2015. 1-14 %arXiv: 1409.1556, 2014. [12] Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation. In: Proceedings of the 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Boston, MA, USA: IEEE, 2015. 3431 -3440 [13] Yuan J Y. Automatic building extraction in aerial scenes using convolutional networks. arXiv: 1602.06564, 2016. [14] Ronneberger O, Fischer P, Brox T. U-Net: Convolutional networks for biomedical image segmentation. In: Proceedings of the 18th Medical Image Computing and Computer-Assisted Intervention. Munich, Germany: Springer, 2015. 234-241 [15] Hariharan B, Arbeláez B, Girshick R, Malik J. Hypercolumns for object segmentation and fine-grained localization. In: Proceedings of the 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Boston, MA, USA: IEEE, 2015. 447-456 [16] Dosovitskiy A, Fischer P, Ilg E, Häusser P, Hazirbas C, Golkov V, et al. FlowNet: Learning optical flow with convolutional networks. In: Proceedings of the 2015 IEEE International Conference on Computer Vision (ICCV). Santiago, Chile: IEEE, 2015. 2758-2766 [17] 陈文康.基于深度学习的农村建筑物遥感影像检测.测绘, 2016, 39(5): 227-230 doi: 10.3969/j.issn.1674-5019.2016.05.010Chen Wen-Kang. Remote sensing image detection of rural buildings based on deep learning algorithm. Surveying and Mapping, 2016, 39(5): 227-230 doi: 10.3969/j.issn.1674-5019.2016.05.010 [18] Silberman N, Sontag D, Fergus R. Instance segmentation of indoor scenes using a coverage loss. In: Proceedings of the 13th European Conference on Computer Vision. Zurich, Switzerland: Springer, 2014. 616-631 [19] Farabet C, Couprie C, Najman L, LeCun Y. Learning hierarchical features for scene labeling. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2013, 35(8): 1915- 1929 doi: 10.1109/TPAMI.2012.231 [20] Zhang A, Liu X M, Gros A, Tiecke T. Building detection from satellite images on a global scale. arXiv: 1707.08952, 2017. [21] Ghaffarian S, Ghaffarian S. Automatic building detection based on supervised classification using high resolution Google earth images. In: Proceedings of the 2014 ISPRS Technical Commission Ⅲ Symposium. Zurich, Switzerland: ISPRS, 2014. 101-106 [22] Shu Z, Hu X Y, Sun J. Center-point-guided proposal generation for detection of small and dense buildings in aerial imagery. IEEE Geoscience and Remote Sensing Letters, 2018, 15(7): 1100-1104 doi: 10.1109/LGRS.2018.2822760 [23] Yang H L, Lunga D, Yuan J Y. Toward country scale building detection with convolutional neural network using aerial images. In: Proceedings of the 2017 IEEE International Geoscience and Remote Sensing Symposium (IGARSS). Fort Worth, TX, USA: IEEE, 2017. 870-873 [24] Sun L, Tang Y Q, Zhang L P. Rural building detection in high-resolution imagery based on a two-stage CNN model. IEEE Geoscience and Remote Sensing Letters, 2017, 14(11): 1998-2002 doi: 10.1109/LGRS.2017.2745900 [25] Vakalopoulou M, Bus N, Karantzalos K, Paragios N. Integrating edge/boundary priors with classification scores for building detection in very high resolution data. In: Proceedings of the 2017 IEEE International Geoscience and Remote Sensing Symposium (IGARSS). Fort Worth, TX, USA: IEEE, 2017. 3309-3312 [26] Canny J. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1986, PAMI-8(6): 679-698 doi: 10.1109/TPAMI.1986.4767851 [27] 卫亚星, 王莉雯.遥感图像增强方法分析.测绘与空间地理信息, 2006, 29(2): 4-7 doi: 10.3969/j.issn.1672-5867.2006.02.002Wei Ya-Xing, Wang Li-Wen. Analysis of enhancement methods about satellite images. Geomatics and Spatial Information Technology, 2006, 29(2): 4-7 doi: 10.3969/j.issn.1672-5867.2006.02.002 [28] Cai B L, Xu X M, Jia K, Qing C M, Tao D C. DehazeNet: An end-to-end system for single image haze removal. IEEE Transactions on Image Processing, 2016, 25(11): 5187-5198 doi: 10.1109/TIP.2016.2598681 [29] Glorot X, Bordes A, Bengio Y. Deep sparse rectifier neural networks. Journal of Machine Learning Research, 2011, 15: 315-323%In: Proceedings of the 14th International Conference on Artificial Intelligence and Statistics. Fort Lauderdale, USA: JMLR W & CP, 2011. 315-323 http://d.old.wanfangdata.com.cn/Periodical/zdhxb201606009 [30] Ioffe S, Szegedy C. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: Proceedings of the 32nd International Conference on Machine Learning. Lille, France: JMLR, 2015. 448-456 [31] Zeiler M D, Krishnan D, Taylor G W, Fergus R. Deconvolutional networks. In: Proceedings of the 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. San Francisco, CA, USA: IEEE, 2010. 2528- 2535 [32] Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 2014, 15(1): 1929-1958 http://d.old.wanfangdata.com.cn/Periodical/kzyjc200606005 [33] Lin T Y, Dollár P, Girshick R, He K M, Hariharan B, Belongie S. Feature pyramid networks for object detection. In: Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, USA: IEEE, 2017. 936-944 [34] Kingma D P, Ba J. Adam: A method for stochastic optimization. In: Proceedings of the 3rd International Conference on Learning Representations. San Diego, CA, USA: 2015. 1-15 期刊类型引用(3)
1. 吕芳芳,楼旭阳,叶倩. 具有死区非线性输入的柔性臂自适应边界控制. 扬州大学学报(自然科学版). 2024(05): 16-24 . 百度学术
2. 谢志勇,朱娟芬,胡小平. 考虑间隙特性的双机械臂模糊自适应鲁棒控制. 现代制造工程. 2022(02): 52-58 . 百度学术
3. 马永浩,张爽,何修宇,刘志杰. 基于连续反演算法的时滞补偿控制综述. 工程科学学报. 2022(06): 1053-1061 . 百度学术
其他类型引用(5)
-