Improved Conditional GAN for Aerial Image Segmentation

Generative Adversarial Network (GAN) is an algorithmic architecture containing two neural networks, placed against each other to generate new synthetic images and it has been used successfully in image segmentation. The paper analyzes different GAN implementations for segmentation of images acquired by aerial robots in a real context of a rural zone in Romania. To improve the segmentation performance, a new GAN network is proposed by adding a new layer. Data augmentation was done by the following techniques: mirroring, rotation, scaling, gray scaling, blurring, sharpening, etc. Five classes of region of interest are considered: floods, vegetations, buildings, roads, and dry land. GAN implementations were tested on CPU, GPU, and TPU, on individual computing devices and in the cloud. A new layer was added. The performances were analyzed in terms of learning time, operating time, and statistical indicators. The batch size was generally low: batches of 1, 4 or 16 images were used in this paper. The results confirm that the use of batch achieves the best training and generalization performance in terms of computational cost, for a wide range of experiments.