基于LeakyMish流行正则化半监督生成对抗网络的图像分类模型

Image Classification Model Based on LeakyMish Manifold Regularization Semi-Supervised Learning with Generative Adversarial Network

在实际应用中,为模型提供大量的人工标签需要消耗大量的人力和财力,因此,近几年基于半监督学习的图像分类问题得到了更多人的关注。半监督生成对抗网络在训练过程中,能够以少量的标签数据训练大量的未标签数据,并取得较好的结果。Improved GAN+Manifold Reg模型是Bruno Lecouat和Chuan-Sheng Foo等人提出的半监督生成对抗网络模型,并且在SVHN数据集和Cifar-10数据集测试中,准确率比知名的ImprovedGAN和TripleGAN高。针对提高图像分类准确率进行研究,在Improved GAN+Manifold Reg模型的基础上进行改进,对激活函数和优化器进行修改,并模仿LeakyReLU激活函数提出LeakyMish激活函数。在SVHN数据集和CIFAR-10数据集测试中,准确率有进一步的提高,而且模型收敛快。

In practical applications,providing a large number of artificial labels for the model requires a lot of human and financial resources.Therefore,in recent years,the problem of image classification based on semi-supervision learning has attracted more attention.During the training process of the semi-supervised learning with generative adversarial network,a large amount of unlabeled data can be trained with a small amount of labeled data,and good results are obtained.The Improved GAN+Manifold Reg model is a semi-supervised learning with generative adversarial network model proposed by Bruno Lecouat and Chuan-Sheng Foo,etc.,and has higher accuracy than the wellknown Improved GAN and Triple GAN in the Svhn dataset and CIFAR-10 dataset tests.The research was aimed at improving the accuracy of image classification.Based on the Improved GAN+Manifold Reg model,the activation function and optimizer were adjusted,and the LeakyMish activation function was proposed,which imitated the LeakyReLU activation function.In the Svhn dataset and CIFAR-10 dataset tests,the accuracy rate had been further improved,and the model converged fast.