针对当训练样本不足时, 传统深度学习算法在手写数字识别中会出现训练不稳定、识别精度较低等问题, 提出了基于条件生成对抗网络的识别方法。首先, 在条件生成对抗网络的基础上, 利用生成器使用类别标签控制图像生成的优点, 将生成器产生的图像样本作为训练数据, 扩充数据集。同时, 利用反卷积网络和卷积网络分别构成生成器和判别器的网络结构, 去掉全连接层以提升模型稳定性。然后, 引入条件批量归一化, 利用它使用类别标签的优点, 使网络学习更多的特征。最后, 改进判别器为分类器, 并提出新的损失函数, 加快收敛速度, 提高识别精度。实验结果表明, 本文所提出的手写数字识别方法生成的图像质量更好, 识别准确率更高, 达到99.43%, 为生成对抗网络及其变体在图像识别领域中的应用提供了参考。

In order to solve the problem of unstable training and low recognition accuracy of traditional deep learning algorithm in handwritten digit recognition when training samples are insufficient, a recognition method based on conditional generative adversarial network is proposed. Firstly, on the basis of the conditional generative adversarial network, the generator of category labels is used to control image generation. The image samples generated by the generator are used to expand the training data set. At the same time, the deconvolution network and the convolutional network are used to construct the network structure of generator and discriminator respectively, and the full connection layer is removed to improve the stability of the model. Then, the conditional batch normalization is introduced. Taking advantage of its use of category labels to make the network learn more features. Finally, the discriminator is improved to be a classifier, and a novel loss function is proposed to speed up the convergence rate and improve the recognition accuracy. The experimental results show that the handwritten digit recognition method proposed in this paper has better image quality and higher recognition accuracy of 99.43%. This paper provides a reference for the application of generative adversarial network and its variants in the field of image recognition.