An Improved WGAN-GP Generative Adversarial Model in View of NIR Spectral 1st Derivative Constraint
LI Zhen-yu1, ZHAO Peng1, 2*
1. College of Computer and Control Engineering, Northeast Forestry University, Harbin 150040, China
2. School of Computer Science and Software Engineering, Guangxi University of Science and Technology, Liuzhou 545006, China
Abstract：The Generative Adversarial Network (GAN) has recently become a hot branch in deep neural networks. The mainstream GAN model consists of many improved versions used in image processing and computer vision. These GAN versions are rarely used in spectral analysis. In spectral analysis, they are mainly used to generate synthetic spectral curves so as to extend the classifier's training set for its augmentation and improve its classification performance. Because of the trend of 1D near infrared (NIR) spectral curve, which is an important classification feature and can be quantitatively denoted by a curve derivation, we improve the current one-class Wasserstein GAN with Gradient Penalty (WGAN-GP) model by imposing a spectral 1st derivative constraint. Specifically, the original NIR spectral vector is connected with the corresponding spectral derivative vector in the revised model L loss function. The concatenated vector is used for model training and spectral curve production. Finally, only the first half is retained in the artificially produced spectral vector to generate the synthetic spectral curves. In our NIR classification experiments of wood species and apple classes, the classification accuracy in some classifiers such as Support Vector Machine (SVM), 1D-Convolutional Neural Network (1D-CNN) and LeNet-5 neural network is increased to some extents after the training set augmentation by use of our improved WGAN-GP model compared with that by use of original WGAN-GP model. Moreover, the NIR spectral curve quality produced by our improved WGAN-GP model has increased greatly, which is indicated by some evaluation measures such as Inception Score, which is computed by use of 1D-CNN instead of the original 2D Inception Net-V3 network, the correlation coefficient between original and synthetic spectral vectors, and these two vectors' difference L1 and L2 norms,compared with that by use of the original WGAN-GP model.
[1] Goodfellow I J, Pouget-Abadie J, Mirza M, et al. Generative Adversarial Nets, Part of Advances in Neural Information Processing Systems 27 (NIPS 2014), 2014.
[2] Kingma D P, Welling M. Auto-Encoding Variational Bayes, 2013, arXiv: 1312. 6114.
[3] Arjovsky M, Chintala S, Bottou L. Wasserstein GAN, 2017, arXiv: 1701. 07875v3.
[4] Berthelot D, Schumm T, Metz L. BEGAN: Boundary Equilibrium Generative Adversariul Networks, 2017, arXiv: 1703. 10717.
[5] Radford A, Metz L, Chintala S. Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks, arXiv: 1511: 06434v2.
[6] Odena A, Olah C, Shlens J. Conditional Image Synthesis with Auxiliary Classifier GANs, 2017, arXiv, 1610. 09585v4.
[7] Wu M, Wang S, Pan S, et al. Scientific Reports, 2021, 11: 23842.
[8] Mishra P, Herrmann I. Chemometrics and Intelligent Laboratory Systems, 2021, 215: 104362.
[9] McHardy R G, Antoniou G, Conn J J A, et al. Analyst, 2023, 148: 3860.
[10] WANG Zhi-hui, DING Li-xia（王志辉，丁丽霞）. Spectroscopy and Spectral Analysis（光谱学与光谱分析）, 2010, 30(7): 1825.
[11] Gulrajani I, Ahmed F, Arjovsky M, et al. Improved Training of Wasserstein GANs, Proceedings of 31st International Conference on Neural Information Processing Systems, 2017, 5769.
[12] Salimans T, Goodfellow I, Zaremba W, et al. Improved Techniques for Training GANs, 30th Conference on Neural Information Processing Systems (NIPS 2016), 2016, 2226.
