Evaluation of Modified FGSM-Based Data Augmentation Method for Convolutional Neural Network-Based Image Classification

Computer vision applications demand a significant amount of data for effective training and inference in many computer vision tasks. However, data insufficiency situations usually happen due to multiple reasons, resulting in computational models whose performance is inadequate. Traditional data augmentation techniques are presented to solve this overfitting problem; however, their application is not always possible or desirable. In this context, this paper addresses a different data augmentation technique for classification methods based on adversarial images to reduce the impact of sample imbalance utilizing the Fast Gradient Sign Method (FGSM) with added noise to enhance classifier performance. To validate the method, a set of images was used for the classification of diseases in coffee plants due to the soil’s lack of nutrients. The results showed an improvement in the model performance for this classification in coffee plants proving the validity of the proposed method, which can be used as an alternative to traditional data augmentation methods.