基于改进TransGAN的零样本图像识别方法

零样本学习算法旨在解决样本极少甚至缺失情况下的图像识别问题。生成式模型通过生成缺失类别的图像，将此问题转化为传统的基于监督学习的图像识别，但生成图像的质量不稳定、容易出现模式崩塌，影响图像识别准确性。为此，通过对TransGAN模型进行改进，提出基于改进TransGAN的零样本图像识别方法。将TransGAN的生成器连接卷积层进行降维，并进一步提取图像特征，使生成图像特征和真实图像特征更加接近，提高特征的稳定性；同时，对判别器加入非线性激活函数，并进行结构简化，使判别器更好地指导生成器，并减小计算量。在公共数据集上的实验结果表明，所提方法的图像识别准确率较基线模型提高了29.02%，且具有较好的泛化性能。

An image recognition method of zero -shot learning based on an improved TransGAN

Zero-shot learning algorithms aim to address the challenge of image recognition with limited or even missing samples. By transforming the problem into a supervised learning task through the use of generative models, the method generates images of missing classes. However, the quality of generated images can be inconsistent and is susceptible to pattern collapse, affecting image recognition accuracy. To address this issue, we propose an improved zero-shot learning image recognition method based on an improved TransGAN. The generator of TransGAN is linked to a convolutional layer for dimensionality reduction, leading to a more effective extraction of image features and improved stability. Moreover, the addition of a nonlinear activation function to the discriminator and simplifying its structure enhances its ability to guide the generator and reduces computational requirements. Experiment results on public datasets show that our proposed method increases image recognition accuracy by 29.02% compared to the baseline model and demonstrates improved generalization performance.