Multiview Generative Adversarial Network and Its Application in Pearl Classification

This paper focuses on automatic pearl classification by adopting deep learning method, using multiview pearl images. Traditionally, in order to get a satisfying classification result, we need to collect a huge number of labeled pearl images, which however is expensive in industry. Fortunately, generative adversarial network (GAN) was proposed recently to effectively expand training set, so as to improve the performance of deep learning models. We thus propose a multiview GAN (MV-GAN) to automatically expand our labeled multiview pearl images, and the expanded data set is then used to train the multistream convolutional neural network (MS-CNN). The experiments show that the utilization of images generated by the MV-GAN can indeed significantly reduce the classification error of the basic MS-CNN (up to 26.71%, relatively), obtaining the state-of-the-art results. More interestingly, it can also help the MS-CNN resist the brightness disturbance, leading to more robust classification.