| 摘 要: 生成对抗网络因为在训练时具有能快速获取真实感、产生大量特征等优点,所以该方法在有监督和半监督的图像识别中逐渐得到广泛应用。文章以提高GAN(Generative Adversarial Networks)模型下的图像识别准确率为目标,基于已有的GAN模型,提出一种基于GAN模型的半监督深度学习模型,并将所建模型放入MNIST、CIFAR-10和Fashion-MNIST三种不同的数据集进行测试,结果显示,SSE-DCGAN模型在三种数据集标签数据较少时,能够很好地识别图像,在三个数据集上识别精度分别达到99.04%、83.66%、89.64%,进行消融实验的结果也表明,在模型中加入编码器后,准确率分别达到0.43%、2.55%、4.44%的提升。 |
| 关键词: 生成对抗网络;半监督;图像识别;特征匹配 |
|
中图分类号: TP391.41
文献标识码: A
|
|
| Research on Deep Convolutional Adversarial Network Model based on Semi-supervised Encoding |
|
OU Lili, DU Fangfang
|
(School of Intelligent Systems Engineering, Huanghe Jiaotong University, Jiaozuo 454950, China)
1270370009@qq.com; 532637120@qq.com
|
| Abstract: Generative adversarial network has been widely used in supervised and semi-supervised image recognition due to its advantages of quickly acquiring realism and generating a large number of features during training. Aiming at improving the accuracy of image recognition under the General Adversarial Networks ( GAN) model, this paper proposes a semisupervised deep learning model based on the existing GAN model, and the proposed model is tested on three different datasets: MNIST, CIFAR-10, and Fashion-MNIST. The results show that the SSE-DCGAN model can effectively recognize images when there is less label data in the three datasets. The recognition accuracy reaches 99.04% , 83.66% , and 89.64% on the three datasets, respectively. The results of ablation experiments also show that after an encoder is added to the model, the accuracy improves by 0.43% , 2.55% , and 4.44% , respectively. |
| Keywords: generative adversarial networks; semi-supervision; image recognition; feature matching |
