PhytoGAN: Unpaired Dead-to-Live Phytoplankton Translation

Shuai Han, Xin Sun, Junyu Dong, Shu Zhang, Qiong Li

Abstract: Detecting phytoplankton that causes red tide is an urgent task. However, the live phytoplankton images are scarce and difficult to obtain. This paper aims to discover a mapping between live and dead phytoplankton. We propose PhytoGAN, a Generative Adversarial Network for the unpaired dead-to-live phytoplankton translation. We design a new PCALoss by extracting the principal component of the image to enhance the contour of the generated image. The addition of PCALoss can significantly improve the integrity of images. The experiments are carried out on the existing phytoplankton dataset. A series of experiments are discussed in the paper to demonstrate the performance of the domain transformation and the proposed loss functions. The experimental results indicate that PhytoGAN can produce more integral images of phytoplankton while completing the domain transformation compared with the existing methods.