Complex conditional generative adversarial nets for multiple objectives detection in aerial images

Abstract

Simultaneously detection and evaluation of small regions of interest from aerial images is successfully achieved by conditional generative adversarial nets (cGAN). As novelty, the paper proposes a cheap and accurate method based on a cGAN structure, containing two generators, and graphics processing units (GPU) to segment small flooded areas and respectively, roads from images taken by unmanned aerial vehicles. In the learning phase the weights for the discriminator and the two generators are established by a back propagation method. The real mask is created by using information of the color components R, G, B, H and a voting scheme in a supervised process. A set of 40 images were used for the learning phase and another set of 100 images were used for method validation. The method presents the advantages of accuracy and time processing (especially in the operational phase).