Improving cyclic GAN for camouflage removal and liver segmentation

Abstract

Deep learning based on machine learning is a representative technology of arti？cial intelligence, which is the main axis of 4th industrial revolution, and its performance is improving in many fields. The development of various structures and learning methods is rapidly improved, and various evolving models of the recently developed Generative Adversarial Network(GAN) algorithms have become the latest research fields to raise the limits of machine learning. The GAN algorithm is a form of unsupervised learning for transforming a noise z vector into a realistic image that fits its purpose through learning on a large amount of data. In this study, we designed two generators for the cyclic GAN learning, and converted the facial images of the combatant facial camouflage into clean face images that can be identified. We propose an iterative learning method that can improve the performance of the network while overcoming the quantitative limit of the image data directly captured from military units. In addition, the experiment of liver and lesion segmentation on liver CT images showed the applied of GAN in medical image processing. The medical image has proposed the network learning method with improved performance by adding the pixel unit accuracy to the objective function for the characteristic of the task.