Improved generative adversarial network with deep metric learning for missing data imputation

Mohammed Al-taezi, Yu Wang, Pengfei Zhu, Qinghua Hu, Abdulrahman Al-Badwi

Abstract: Incomplete data are ubiquitous in real-world computer vision tasks. Imputing missing data is crucial for modeling machine learning algorithms. Although existing methods, such as MisGAN, have achieved considerable success, two problems remain: (1) generating a complete image from random noise weakens the model’s imputation capability; (2) owing to the influence of random noise, the training of the generator is unstable. Therefore, this study proposes a novel approach that uses deep metric learning to enable MisGAN to perform multi-tasking missing data imputation. First, an image feature extraction network is applied to extract the semantic representation of the images. Then, deep metric learning is performed to learn good features embedding by maximizing inter-class variation and minimizing intra-class variation. Such an embedding replaces the random noise and is fed into MisGAN to obtain an improved generated complete image. Several experiments are conducted on datasets, demonstrating that the proposed method can significantly outperform baseline methods.