基于谱约束的生成对抗网络图像数据生成研究

数据匮乏是深度学习面临的一大难题。利用生成对抗网络(GAN)能够基于语义生成新的图像数据这一特性,提出一种基于谱约束的生成对抗网络图像数据生成方法,该方法针对卷积生成对抗网络模型易崩溃不收敛的问题,从每层神经网络的参数矩阵W的谱范数角度出发,引入谱范数归一化网络参数矩阵,将网络梯度限制在固定范围内,减缓判别网络收敛速度,从而提高GAN的训练稳定性。实验表明,通过该方法生成的数据相比原始GAN以及DCGAN、WGAN等生成的图像样本数据在图像识别网络中具有更高的准确率,能够对少量样本数据进行有效扩充。

Image Data Generation Based on Generative Adversarial Network

Insufficient data is a big problem for deep learning. In this paper, based on the feature that the generative adversarial networks (GAN) can semantically generate new data, a new GAN image data generation method based on spectral restriction was proposed. For the problems of collapse and divergence of deep convolution generation adversarial networks (DCGAN), this proposed method starts from the spectral norm of the parameter matrix W of each layer of the neural network, and introduces the spectral norm to normalize the network parameter matrix, thus the network gradient is limited to a fixed range, which slows down the discriminative network convergence speed and improves the training stability of GAN. Experimental results show that in the image recognition network, the proposed method presents higher accuracy than original GAN, DCGAN, WGAN and other methods, expanding a small amount of sample data efficiently.