Image classification based on principal component analysis optimized generative adversarial networks
- PDF / 1,817,178 Bytes
- 15 Pages / 439.642 x 666.49 pts Page_size
- 111 Downloads / 223 Views
Image classification based on principal component analysis optimized generative adversarial networks Chunzhi Wang1 · Pan Wu1 · Lingyu Yan1
· Zhiwei Ye1 · Hongwei Chen1 · Hefei Ling2
Received: 3 June 2020 / Revised: 15 September 2020 / Accepted: 23 October 2020 / © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Recently, the generative adversarial networks(GAN) has been widely used in various fields of machine learning. It avoids the complicated solving process of the original generation model while ensuring the generation effect. However, since the inputs of GAN are random initialized, it takes a long time to train the data generated by the model to fit the original data distribution. Therefore, in this paper, we propose a principal component analysis optimized generative adversarial networks (PCA-GAN). The original data is compressed and reduced by principal component analysis to generate the input of the confrontation network, so that the input data retains the characteristics of the original data to some extent, thereby improving the data generation performance and reducing the training time cost. We applied our PCA-GAN to image classification, and the experimental results show that the model effectively improve the accuracy of image classification and enhance the stability of the model. Keywords Generative adversarial networks · Image classification · Principal component analysis · Semi-supervised learning Lingyu Yan
[email protected] Chunzhi Wang [email protected] Pan Wu [email protected] Zhiwei Ye [email protected] Hongwei Chen [email protected] Hefei Ling hefei [email protected] 1
School of Computer Science, Hubei University of Technology, Wuhan, China
2
School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China
Multimedia Tools and Applications
1 Introduction Recently, due to the good performance of machine learning algorithm [16–19], several image classification methods have been proposed.The generative adversarial networks(GAN) [5] is a novel machine learning structural model proposed by Professor Ian Goodfellow of the University of Montreal in 2014, which has achieved good results in data generation. Since GAN could generate random samples which are similar to the real data distribution through model training and learning, More and more scholars are engaged in the research of GAN [16]. Arjovsky proposed an improvement to GAN in the measurement of data similarity, namely Wasserstein GAN [2]. He used Wasserstein distance to replace the distance measurement formula of the probability distribution in the original GAN, and optimized the instability and model collapse of the original GAN training process. EBGAN [20] improved GAN from the perspective of energy model, which gives GAN a definition of an energy model that is different from other models in defining loss functions by distance metrics, training the GAN model with a broader structure and a wider variety of loss function types. DCGAN (Deep Convolutional Generative
Data Loading...