Optimized Class-specific Data Augmentation for Plant Stress Classification

Nasla Saleem, Baskar Ganapathysubramanian, Aditya Balu, Talukder Zaki Jubery, Soumik Sarkar, Arti Singh, Asheesh K Singh

Keywords: Plant stress classification, Deep Learning, Automated Data Augmentation

TL;DR: This paper study explores an automated data augmentation workflow for training deep learning models that are optimized by GA

Abstract:

Data augmentation has the potential to significantly improve the performance of deep learning-based image classifiers. However, a key challenge in applying data augmentation is choosing an effective set of augmentations from a large pool of candidates. Recently, automated augmentation strategies have produced state-of-the-art results for image classification. Most results have focused on improving the total accuracy of the classifier, often at the cost of reduced performance of a finite number of classes. We explore a Genetic Algorithm-based optimization to identify the ideal class-specific augmentations that maximize the mean-per-class accuracy, starting from a well-trained classifier (which serves as our baseline). We illustrate the utility of this strategy on a well-studied problem (and
associated dataset) of classifying soybean leaf stresses. Our (preliminary) work indicated improvements over our baseline model and showed improvement in the mean-per-class accuracy from 90.68% to 93.11% across generations. Identifying class-specific augmentations can provide contextual information to end users. This approach is computationally less expensive than traditional Network-Architecture-Search (NAS), as we only seek to fine-tune the baseline classifier.