We know deep learning is one of the main techniques to achieve state-of-the-art performances in computer vision methods applied to agriculture. However, this technique requires very large datasets (thousands of images) so that the automatic feature extraction can work accurately, generalize and not overfit. What can be done if we have a small dataset? There are several possibilities, but one of the simplest is data augmentation.
Data augmentation consists basically of creating variations of an original image by rotating from various angles, cropping, scaling, blurring, noising, gray-scaling, or mirroring. This way, the essence of the classes we are interested in is reinforced because the network has to learn how to extract the permanent features that really define the object of interest. Thus, the chances of overfitting are also reduced and the gap between training and test performances is lower.
But there is more. In recent years, some libraries have also implemented the simulation of specific climate conditions, such as rain, snow, or fog. For instance, Albumentations is a Python library for creating different image augmentation pipelines. It implements a rich variety of image transform operations for different computer vision tasks, including object classification, segmentation, and detection.
In some of our notebooks (https://edenlibrary.ai/notebooks), we deeply cover the Data Augmentation technique. They all are working examples ready for your modifications and you check them by clicking the link below.
If you find this repository interesting, please do not forget to give us a star and share it with your colleagues.