一种融合生成对抗网络的零样本图像分类方法

A Zero-shot Classification Based on Generative Adversarial Network

目前很多零样本图像分类方法是采用学习语义信息空间和图像特征空间之间的映射关系来实现图像分类,但是这些方法会产生枢纽化和域漂移问题。通过利用生成对抗网络实现图像特征生成可以缓解以上两种问题,但该方法容易产生模式崩溃从而导致生成的图像特征不真实。因此提出一种改进的生成对抗网络方法,通过在生成器网络上增加一个重构网络,将生成器生成的图像特征重构回语义信息,以此实现生成器网络生成的图像特征更加符合语义信息的图像特征。实验结果表明,该方法相较于原本的生成对抗网络模型而言,在AWA、CUB、FLO、SUN四个数据集上的分类准确率分别提升了1.0、0.1、1.2和0.9个百分点,证明了通过融合改进的生成对抗网络实现零样本图像分类方法的有效性和可行性。

Since most zero-shot image classification adopts the mapping relationship between the semantic information space and the image feature space to achieve the image classification,which will cause hubness and domain drift.The above two problems can be alleviated by using the generated adversarial network to generate the image feature,but this method is prone to mode collapse,which will result in unrealistic image features.Therefore,an improved generative adversarial network method is proposed.By adding a reconstruction network to the generator network,the image features generated by the generator are reconstructed back to semantic information,so that it is more in line with the semantic information.The experiment shows that compared with the original generative adversarial network model,the proposed method has increased by 1.0,0.1,1.2,and 0.9 percent in classification accuracy on the four datasets of AWA,CUB,FLO,and SUN,respectively.It proves the effectiveness and feasibility of the zero-shot image classification method by fusing the improved generative adversarial network.