Agricultural Image Augmentation with Generative Adversarial Networks GANs

Deep Learning gains popularity in almost every field of research currently and agricultural industry is not the exception in this. One of the main challenge in deep learning is the requirement of lots of data in order to produce good accuracy and results. But the main problem in agricultural field is the lack of dataset, more specifically image-based dataset. It takes more time and resources than any other to collect data in this field. This paper proposed a method by which one can generate multiple yet different images and also original look like images from very low amount of image datasets. Then this enlarged dataset can be used to train any deep learning network more effectively. We have used very simple method of Generative Adversarial Networks (GANs) to expand the dataset. For experiment, a dataset of only 52 images of rice leaves infected with brown-spot disease has been considered here. We have successfully generated original look like but fake samples of the original dataset which are selected to extend the dataset. The generator model has been designed in very user friendly way. It only needs the number of images that a user wants to generate and the quality of the images as user inputs. The model will calculate rest of the necessary parameters by its own and generate the samples. The quality of the generated images is validated with the classification accuracy in VGG16 deep learning network. The extended dataset with the generated images achieved 98% accuracy in this case which is 1% better than original one.