A Method of Image Classification for Objects with Camouflaged Color Features
-
摘要: 针对图像中某几类物体具有相似颜色特征而导致的分类困难问题,本文提出了一种具有隐蔽色特征物体的图像分类方法。该方法针对可见光图像中具有颜色隐蔽性物体而难以区分的问题,通过将二维图像的邻域像素空间特征与高光谱图像的谱段特征相结合并使用改进的局部线性嵌入降维算法实现了空谱联合的特征降维,最终利用主动学习胶囊网络训练高光谱数据分类器从而实现场景内目标的分类。通过改进的主动学习函数可以对更具代表性的样本进行标注,实现了利用小样本集对胶囊网络的训练,有效降低了样本的标注成本和模型的训练成本,提高了模型分类性能。实验表明,该算法运行在自建高光谱数据集上能够有效地分类隐蔽色特征物体和其他自然场景,针对隐蔽色目标的平均准确率达到了91%,针对所有类别物体的平均准确率达到了89.9%。Abstract: To solve the classification difficulty caused by different objects with similar color features in an image, this paper proposes an image classification method for objects with camouflaged color features. To alleviate the difficulty of distinguishing color-camouflaged objects in RGB images, this method not only combines the spatial domain features of the neighborhood pixels and the spectral domain features of the hyperspectral data, which realizes the spatial-spectrum joint feature construction, but also uses the improved LLE(local linear embedding) algorithm to accomplish spectral dimensional reduction. The proposed method uses an active learning capsule network to train a hyperspectral data classifier and classifies objects in the scene. Active learning can label more representative samples through the improved active learning function and realized capsule network training based on a minor sample dataset, which reduces the cost of sample labeling and model training significantly, thereby improving the classification performance of the model. Experiments show that the algorithm proposed in this paper can effectively classify camouflaged targets and other natural targets based on our self-made hyperspectral dataset. The average accuracy of camouflaged targets was 91%, and the average accuracy of all target types was 89.9%.
-
表 1 不同标注方法的分类性能比较
Table 1. Comparison of detection performance of differentmethods of labelling
Category Percentage 0% 10% 20% 30% 40% Camouflage target – car 87.2% 87.3 88.2 % 90.8% 90.3% Camouflage target – net 88.1% 88.1% 89.7% 91.2% 91.4% Grass 86.9% 87.4% 88.1% 88.7% 88.2% Trees 87.9% 87.2% 86.3% 91.1% 91.3% Shadow 83.4% 84.0% 85.2% 86.3% 86.4% Sky 86.8% 88.5% 90.8% 91.0% 91.2% Average 86.7% 86.8% 88.1% 89.9% 89.8% 表 2 不同分类方法的性能比较
Table 2. Comparison of detection performance of differentmethods of classification
Category Method LR AlexNet Capsnet Camouflage target – car 88.2% 87.6% 90.8% Camouflage target – net 87.9% 87.1% 91.2% Grass 86.3% 84.9% 88.7% Trees 87.4% 88.5% 91.1% Shadow 86.2% 82.3% 86.3% Sky 89.6% 87.9% 91.0% Average 87.6% 86.4% 89.9% -
[1] SONG W, LI S, FANG L, et al. Hyperspectral image classification with deep feature fusion network[J]. IEEE Transactions on Geoscience and Remote Sensing, 2018, 99: 1-12. http://ieeexplore.ieee.org/document/8283837/citations [2] 罗建华, 李明奇, 郑泽忠, 等. 基于深度卷积神经网络的高光谱遥感图像分类[J]. 西华大学学报: 自然科学版, 2017, 36(4): 13. https://www.cnki.com.cn/Article/CJFDTOTAL-SCGX201704003.htmLUO Jianhua, LI Mingqi, ZHENG Zezhong, et al. Hyperspectral remote sensing images classification using a deep convolutional neural network model[J]. Journal of Xihua University: Natural Science Edition, 2017, 36(4): 13. https://www.cnki.com.cn/Article/CJFDTOTAL-SCGX201704003.htm [3] CHEN Y, ZHAO X, JIA X. Spectral–spatial classification of hyperspectral data based on deep belief network[J]. IEEE Journal of Selected Topics in Applied Earth Observations & Remote Sensing, 2015, 8(6): 2381-2392. http://ieeexplore.ieee.org/document/7018910/ [4] ZHAO W, DU S. Learning multiscale and deep representations for classifying remotely sensed imagery[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2016, 113: 155-165. doi: 10.1016/j.isprsjprs.2016.01.004 [5] 张倩, 董安国, 宋睿. 基于多特征和改进自编码的高光谱图像分类[J]. 激光与光电子学进展, 2020, 57(8): 081010. https://www.cnki.com.cn/Article/CJFDTOTAL-JGDJ202008012.htmZHANG Qian, DONG Anguo, SONG Rui. Hyperspectral image classification based on multiple features and an improved autoencoder[J]. Laser & Optoelectronics Progress, 2020, 57(8): 081010. https://www.cnki.com.cn/Article/CJFDTOTAL-JGDJ202008012.htm [6] 陶志勇, 李杰, 唐晓亮. 融合小波变换与胶囊网络的纹理图像分类算法[J]. 激光与光电子学进展, 2020, 57(24): 241002. https://www.cnki.com.cn/Article/CJFDTOTAL-JGDJ202024006.htmTAO Zhiyong, LI Jie, TANG Xiaoliang. Texture image classification algorithm combining wavelet transform and capsule network[J]. Laser & Optoelectronics Progress, 2020, 57(24): 241002. https://www.cnki.com.cn/Article/CJFDTOTAL-JGDJ202024006.htm [7] Sabour S, Frosst N, Hinton G E. Dynamic routing between capsules[C]//Advances in Neural Information Processing Systems, 2017: 3856-3866. [8] Cohn D A, Ghahramani Z, Jordan M I. Active learning with statistical models[J]. Journal of Artificial Intelligence Research, 1996, 4(1): 705-712. http://arxiv.org/abs/cs/9603104v1 [9] Tuia D. Unbiased query-by-bagging active learning for VHR image classification[C]//Proceedings of SPIE, 2010, 7830: 78300K. [10] Quadrianto N, Petterson J, Smola A J. Distribution matching for transduction[C]//Advances in Neural Information Processing Systems, 2009: 1500-1508. [11] 纪磊, 张欣, 张丽梅, 等. 基于空谱加权近邻的高光谱图像分类算法[J]. 激光与光电子学进展, 2020, 57(6): 061013. https://www.cnki.com.cn/Article/CJFDTOTAL-JGDJ202006017.htmJI Lei, ZHANG Xin, ZHANG Limei, et al. Hyperspectral image classification algorithm based on space-spectral weighted nearest neighbor[J]. Laser & Optoelectronics Progress, 2020, 57(6): 061013. https://www.cnki.com.cn/Article/CJFDTOTAL-JGDJ202006017.htm [12] WANG J. Locally Linear Embedding[M]// Geometric Structure of High -Dimensional Data and Dimensionality Reduction, Springer Berlin Heidelberg, 2012: 203-220.