GAN

GAN is considered as one of the most useful neural networks in many fields. Mainly GAN is used to find the feature loss in image processing caused by down sampling. When the image is compressed, some of the information may get lost or quality of that image is lost, so we may need to recover all the original details. For this recovery, a perpetual loss function comprising adversial loss and content loss is defined. This function is then compared with the widely used pixel-​wise mean squared error (MSE) loss. While working on a large number of images, this model is able to improve the quality of highly compressed images. This becomes important with all models which contain image processing work, mainly agriculture, because certain applications are dependent on remote sensing images.

Barth et al. (2017) proposed a model to overcome the problems associated with big amount of data obtained in deep learning systems. In the absence of manually marked information a large quantity of data (as in deep learning model or GAN-​based model) is used. This is called unsupervised cycle, or generative adversarial system, to optimize the practicality of artificial agricultural pictures. Authors have proposed 10,500 artificial, 50 empirically annotated and 225 unlabelled empirical pictures to get their model working. The hypothesis made was that there was resemblance between synthetic images and empirical images which can be enhanced qualitatively to improve the transformation of features. Because of this analysis the artificial pictures were transformed easily on local characteristics like light diffusion, colour and consistency as compared with global feature translation, which was not that good.