RL-NST: Augmented Corn Leaf Disease Dataset

The complete dataset is available in the RL-NST repository. It consists of 4,776 augmented corn leaf disease images generated using a Reinforcement Learning–based Neural Style Transfer (RL-NST) framework. This dataset extends existing resources, including PlantVillage, CCMT, and field-collected samples. It comprises five classes: Common Rust (960 images), Leaf Blight (783 images), Leaf Spot (944 images), Streak Virus (1,890 images), and Healthy (199 images). All images are provided in JPEG format with standardized resolution, making them suitable for training and benchmarking deep learning models for plant disease detection.

The RL-NST implementation code is also available in the repository. It includes the complete framework described in the manuscript: Kanchanadevi K, Sandhia G K. Adaptive Image Augmentation Using Reinforcement Learning and Neural Style Transfer for Corn Disease Diagnosis. PeerJ Computer Science (under review).

The proposed framework integrates Neural Style Transfer (NST) with Reinforcement Learning (RL) to generate diverse and high-quality augmented images, thereby improving classification performance in corn leaf disease diagnosis.

Key features of the repository include:

A custom NST model with residual connections for enhanced style transfer quality

A DQN-based RL agent for dynamic optimization of NST hyperparameters

Automated generation of augmented images to balance datasets and reduce overfitting

Support for transfer learning evaluation using AlexNet, VGG19Net, and InceptionV3Net

The repository is structured into modules for model implementation, data handling, training, and result storage. It also provides scripts for training the RL-NST model, preprocessing data, and evaluating performance using metrics such as classification accuracy, diversity score, content loss, style loss, and Structural Similarity Index (SSI).

Overall, the RL-NST framework generates high-quality augmented images that improve both classification accuracy and generalization of deep learning models while preserving essential content and style characteristics.