Enhancing Performance of Deep Learning Models with a Novel Data Augmentation Approach

Data augmentation is a popular deep learning strategy for increasing the size of the training set and improving the generalisation capacity of deep learning models, particularly in computer vision problems. This paper provides an overview of popular data augmentation techniques used in deep learning models (like CNNs), such as rotation, flipping, cropping, scaling, neural style transfer, grayscale and adding noise to the input images. We propose a new data augmentation technique which involves four sequential steps: rescaling, rotating, flipping, and converting to grayscale to enhance the performance of CNNs further. Using a dataset with numerous parameters, we show that the suggested strategy outperforms state-of-the-art data augmentation methods. The accuracy of the model trained using the suggested technique was 93.33%, representing a 3.33% improvement over the accuracy of the model built using typical data augmentation techniques.