Underwater Optical Image Recognition Based on Dual Flexible Metric Adaptive Weighted 2DPCA
-
摘要: 受观测条件和采集场景等因素影响,水下光学图像通常呈现出高维小样本特性且易伴随着噪声信息干扰,导致许多降维方法对其识别过程中的鲁棒表现力不足。为解决上述问题,该文提出一种新颖的双灵活度量自适应加权2维主成分分析方法(DFMAW-2DPCA)应用于水下图像识别。该方法不仅在建立重构误差和方差之间双层关系中同时使用了灵活的鲁棒距离度量机制,而且能够根据每个样本实际状态自适应学习到与之相匹配的权重,有效增强了模型在水下噪声干扰环境下的鲁棒性并实现识别精度的提升。与此同时,该文设计了一个快速非贪婪算法用于最优解的获取,其具有良好的收敛性。通过3个水下图像数据库中进行大量实验的结果表明,DFMAW-2DPCA在同类方法中具有更为杰出的整体性能。Abstract: Influenced by factors such as observation conditions and acquisition scenarios, underwater optical image data usually presents the characteristics of high-dimensional small samples and is easily accompanied with noise interference, resulting in many dimension reduction methods lacking robust performance in their recognition process. To solve this problem, a novel 2DPCA method for underwater image recognition, called Dual Flexible Metric Adaptive Weighted 2DPCA (DFMAW-2DPCA), is proposed. DFMAW-2DPCA not only utilizes a flexible robust distance metric mechanism in establishing a dual-layer relationship between reconstruction error and variance, but also adaptively learn matching weights based on the actual state of each sample, which effectively enhances the robustness of the model in underwater noise interference environments and improves recognition accuracy. In this paper, a fast nongreedy algorithm for obtaining the optimal solution is designed and has good convergence. The extensive experimental results on three underwater image databases show that DFMAW-2DPCA has more outstanding overall performance than other 2DPCA-based methods.
-
1 DFMAW-2DPCA优化求解算法
输入:样本增广矩阵$ {\boldsymbol{A}} \in {\Re ^{mN \times n}} $,特征维度$ k $和$ p \in \left( {0,2} \right) $,其中样本数据$ {{\boldsymbol{A}}_i} $已完成数据中心化处理。 初始化:$ {{\boldsymbol{V}}^{\left( {t - 1} \right)}} \in {{\boldsymbol{R}}^{n \times k}} $,其满足$ {{\boldsymbol{V}}^{\mathrm{T}}}{\boldsymbol{V}} = {{\boldsymbol{I}}_k} $,$ t = 1 $,$ \delta = 0.01 $。 当 不收敛时 执行 1:分别利用式(8)、式(10)和式(11)计算对角矩阵$ {{\boldsymbol{D}}^{\left( {t - 1} \right)}} $,$ {{\boldsymbol{S}}^{\left( {t - 1} \right)}} $和$ {{\boldsymbol{G}}^{\left( {t - 1} \right)}} $的对角元素$ \alpha _{ij}^{\left( {t - 1} \right)} $,$ s_{ij}^{\left( {t - 1} \right)} $和$ g_{ij}^{\left( {t - 1} \right)} $。 2:利用式(16)计算$ l_{ij}^{\left( {t - 1} \right)} $,并同时构建由对角元素$ {1 \mathord{\left/ {\vphantom {1 {l_{ij}^{\left( {t - 1} \right)}}}} \right. } {l_{ij}^{\left( {t - 1} \right)}}} $所组成的对角矩阵$ {L^{\left( {t - 1} \right)}} $。 3:计算对角矩阵$ {{\boldsymbol{U}}^{\left( {t - 1} \right)}} $的对角元素$ u_{ij}^{\left( {t - 1} \right)} $,其中$ u_{ij}^{\left( {t - 1} \right)} = {1 \mathord{\left/ {\vphantom {1 {l_{ij}^{\left( {t - 1} \right)}}}} \right. } {l_{ij}^{\left( {t - 1} \right)}}}\left( {s_{ij}^{\left( {t - 1} \right)} + \alpha _{ij}^{\left( {t - 1} \right)}g_{ij}^{\left( {t - 1} \right)}} \right) $。 4:计算加权协方差矩阵$ {{\boldsymbol{A}}^{\mathrm{T}}}{{\boldsymbol{U}}^{\left( {t - 1} \right)}}{\boldsymbol{A}} $。 5:求解目标函数(8)的最优投影矩阵$ {{\boldsymbol{V}}^{\left( t \right)}} $,$ {{\boldsymbol{V}}^{\left( t \right)}} $是由$ {{\boldsymbol{A}}^{\mathrm{T}}}{{\boldsymbol{U}}^{\left( {t - 1} \right)}}{\boldsymbol{A}} $的前$ k $个最大特征值所对应特征向量组成。 6:检验收敛条件$ J({{\boldsymbol{V}}^{\left( t \right)}}) - J({{\boldsymbol{V}}^{\left( {t - 1} \right)}}) \le \delta $满足;如果满足,结束循环;否则执行步骤7。 7:通过获取到的$ {{\boldsymbol{V}}^{\left( t \right)}} $完成对角矩阵$ {{\boldsymbol{D}}^{\left( t \right)}} $,$ {{\boldsymbol{S}}^{\left( t \right)}} $和$ {{\boldsymbol{G}}^{\left( t \right)}} $中的每个对角元素$ \alpha _{ij}^{\left( t \right)} $,$ s_{ij}^{\left( t \right)} $和$ g_{ij}^{\left( t \right)} $的计算。 8:根据$ {{\boldsymbol{V}}^{\left( t \right)}} $,$ \alpha _{ij}^{\left( t \right)} $,$ s_{ij}^{\left( t \right)} $和$ g_{ij}^{\left( t \right)} $执行对于对角矩阵$ {{\boldsymbol{L}}^{\left( t \right)}} $中每个对角元素$ {1 \mathord{\left/ {\vphantom {1 {l_{ij}^{\left( t \right)}}}} \right. } {l_{ij}^{\left( t \right)}}} $的计算。 9:完成对角矩阵$ {{\boldsymbol{U}}^{\left( t \right)}} $中每个对角元素$ u_{ij}^{\left( t \right)} $的计算。 10:$ t \leftarrow t + 1 $。 结束循环 输出:$ {{\boldsymbol{V}}^{\left( t \right)}} \in {{\boldsymbol{R}}^{n \times k}} $。 表 1 NF数据库中每种方法的平均最优识别准确率(%)和平均最小重构误差及其所对应的标准差
2DPCA-L1 F-2DPCA Angle-2DPCA GC-2DPCA Cos-2DPCA 2DPCA-2-LP DFMAW-2DPCA p =0.5 p = 1 p =1.5 识别精度 80.25
±0.7685.77
±0.6987.16
±0.8287.50
±0.8588.27
±0.6688.64
±0.6589.85
±0.6088.42
±0.6889.38
±0.64重构误差 462.47
±2.14415.86
±1.92391.25
±2.01382.04
±1.97364.59
±1.88356.78
±1.90326.14
±1.81357.51
±1.92338.92
±1.86表 2 JEDI数据库中每种方法的平均最优识别准确率(%)和平均最小重构误差及其所对应的标准差
2DPCA-L1 F-2DPCA Angle-2DPCA GC-2DPCA Cos-2DPCA 2DPCA-2-LP DFMAW-2DPCA p =0.5 p = 1 p =1.5 识别精度 68.70±0.64 73.15±0.72 73.77±0.73 74.29±0.67 75.06±0.58 75.63±0.61 76.67±0.56 75.30±0.58 76.07±0.53 重构误差 226.59±1.75 210.36±1.68 193.42±1.84 178.90±1.65 155.81±1.60 146.04±1.57 121.53±1.62 149.64±1.67 133.42±1.63 表 3 EPIDHEU数据库中每种方法的平均最优识别准确率(%)和平均最小重构误差及其所对应的标准差
2DPCA-L1 F-2DPCA Angle-2DPCA GC-2DPCA Cos-2DPCA 2DPCA-2-LP DFMAW-2DPCA p =0.5 p = 1 p =1.5 识别精度 77.38±1.57 82.79±1.81 83.71±1.73 84.25±1.59 84.96±1.55 85.50±1.52 86.04±1.54 85.25±1.50 86.54±1.56 重构误差 300.47±3.21 279.80±3.08 268.92±3.03 261.76±3.10 245.28±2.93 236.04±2.97 232.42±2.90 240.53±2.89 227.34±2.92 表 4 EPIDHEU数据库中部分示例样本的可视化识别结果
表 5 3个水下图像数据库中每种方法的平均运行时间与对应的标准差
2DPCA-L1 F-2DPCA Angle-2DPCA GC-2DPCA Cos-2DPCA 2DPCA-2-LP DFMAW-2DPCA NF 11.26±0.58 2.13±0.15 5.60±0.44 4.79±0.37 3.08±0.30 3.15±0.34 3.19±0.25 JEDI 7.53±0.46 1.84±0.12 3.37±0.49 2.68±0.34 2.26±0.32 2.31±0.28 2.37±0.21 EPIDHEU 6.24±0.77 1.52±0.18 2.64±0.61 2.21±0.40 1.80±0.39 1.83±0.36 1.88±0.32 -
[1] ZHOU Jingchun, LIU Qian, JIANG Qiuping, et al. Underwater camera: Improving visual perception via adaptive dark pixel prior and color correction[J]. International Journal of Computer Vision, 2023, 72(2): 1574–1585. doi: 10.1007/s11263-023-01853-3. [2] 郭银景, 吴琪, 苑娇娇, 等. 水下光学图像处理研究进展[J]. 电子与信息学报, 2021, 43(2): 426–435. doi: 10.11999/JEIT190803.GUO Yinjing, WU Qi, YUAN Jiaojiao, et al. Research progress on underwater optical image processing[J]. Journal of Electronics & Information Technology, 2021, 43(2): 426–435. doi: 10.11999/JEIT190803. [3] JOLLIFFE I T and CADIMA J. Principal component analysis: A review and recent developments[J]. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 2016, 374(2065): 20150202. doi: 10.1098/rsta.2015.0202. [4] MARTINEZ A M and KAK A C. PCA versus LDA[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2001, 23(2): 228–233. doi: 10.1109/34.908974. [5] HE Xiaofei, YAN Shuicheng, HU Yuxiao, et al. Face recognition using Laplacianfaces[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2005, 27(3): 328–340. doi: 10.1109/TPAMI.2005.55. [6] YANG Jian, ZHANG D, FRANGI A F, et al. Two-dimensional PCA: A new approach to appearance-based face representation and recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2004, 26(1): 131–137. doi: 10.1109/TPAMI.2004.1261097. [7] LI Ming and YUAN Baozong. 2D-LDA: A statistical linear discriminant analysis for image matrix[J]. Pattern Recognition Letters, 2005, 26(5): 527–532. doi: 10.1016/j.patrec.2004.09.007. [8] CHEN Sibao, ZHAO Haifeng, KONG Min, et al. 2D-LPP: A two-dimensional extension of locality preserving projections[J]. Neurocomputing, 2007, 70(4/6): 912–921. doi: 10.1016/j.neucom.2006.10.032. [9] WANG Rong, NIE Feiping, YANG Xiaojun, et al. Robust 2DPCA with non-greedy l1-norm maximization for image analysis[J]. IEEE Transactions on Cybernetics, 2015, 45(5): 1108–1112. doi: 10.1109/TCYB.2014.2341575. [10] WANG Haixian and WANG Jing. 2DPCA with L1-norm for simultaneously robust and sparse modelling[J]. Neural Networks, 2013, 46: 190–198. doi: 10.1016/j.neunet.2013.06.002. [11] WANG Jing. Generalized 2-D principal component analysis by Lp-norm for image analysis[J]. IEEE Transactions on Cybernetics, 2016, 46(3): 792–803. doi: 10.1109/TCYB.2015.2416274. [12] 王肖锋, 陆程昊, 郦金祥, 等. 广义余弦二维主成分分析[J]. 自动化学报, 2022, 48(11): 2836–2851. doi: 10.16383/j.aas.c190392.WANG Xiaofeng, LU Chenghao, LI Jinxiang, et al. Generalized cosine two-dimensional principal component analysis[J]. Acta Automatica Sinica, 2022, 48(11): 2836–2851. doi: 10.16383/j.aas.c190392. [13] LI Tao, LI Mengyuan, GAO Quanxue, et al. F-norm distance metric based robust 2DPCA and face recognition[J]. Neural Networks, 2017, 94: 204–211. doi: 10.1016/j.neunet.2017.07.011. [14] GAO Quanxue, XU Sai, CHEN Fang, et al. R1-2-DPCA and face recognition[J]. IEEE Transactions on Cybernetics, 2019, 49(4): 1212–1223. doi: 10.1109/TCYB.2018.2796642. [15] ZHAO Meixiang, JIA Zhigang, CAI Yunfeng, et al. Advanced variations of two-dimensional principal component analysis for face recognition[J]. Neurocomputing, 2021, 452: 653–664. doi: 10.1016/j.neucom.2020.08.083. [16] ZHOU Gongyu, XU Guangquan, HAO Jianye, et al. Generalized centered 2-D principal component analysis[J]. IEEE Transactions on Cybernetics, 2021, 51(3): 1666–1677. doi: 10.1109/TCYB.2019.2931957. [17] GAO Quanxue, MA Lan, LIU Yang, et al. Angle 2DPCA: A new formulation for 2DPCA[J]. IEEE Transactions on Cybernetics, 2018, 48(5): 1672–1678. doi: 10.1109/TCYB.2017.2712740. [18] WANG Xiaofeng, SHI Leyan, LIU Jun, et al. Cosine 2DPCA with weighted projection maximization[J]. IEEE Transactions on Neural Networks and Learning Systems, 2023, 34(12): 9643–9656. doi: 10.1109/TNNLS.2022.3159011. [19] ZHANG Huanxing, BI Hongxu, WANG Xiaofeng, et al. A joint-norm distance metric 2DPCA for robust dimensionality reduction[J]. Information Sciences, 2023, 640: 119036. doi: 10.1016/j.ins.2023.119036. [20] HOLMES T H, WILSON S K, VANDERKLIFT M, et al. The role of Thalassoma lunare as a predator of juvenile fish on a sub-tropical coral reef[J]. Coral Reefs, 2012, 31(4): 1113–1123. doi: 10.1007/s00338-012-0934-8. [21] JIA Tianlong, KAPELAN Z, DE VRIES R, et al. Deep learning for detecting macroplastic litter in water bodies: A review[J]. Water Research, 2023, 231: 119632. doi: 10.1016/j.watres.2023.119632. [22] XU Jian, BI Pengfei, DU Xue, et al. Robust PCANet on target recognition via the UUV optical vision system[J]. Optik, 2019, 181: 588–597. doi: 10.1016/j.ijleo.2018.12.098.