Enhancing Strawberry Disease and Quality Detection: Integrating Vision Transformers with Blender-Enhanced Synthetic Data and SwinUNet Segmentation Techniques
Abstract
Agricultural productivity in strawberry cultivation was enhanced through the application of machine learning in this study. Traditional methods for detecting diseases and assessing ripeness in strawberries were identified as labor-intensive and error-prone, which limited farming efficiency and reduced crop yields. To address these challenges, it was hypothesized that advanced machine learning models incorporating attention mechanisms could significantly improve these tasks. The objective was to evaluate the effectiveness of various models by optimizing them for specific agricultural applications. Two datasets of strawberry images were augmented, and three pretrained models—Vision Transformer (ViT), MobileNetV2, and ResNet18—were fine-tuned. Data quality was improved through background removal and noise reduction, and weighted training was employed to manage imbalanced class distributions. The robustness of the models was further tested using synthetic data generated via Blender to simulate data scarcity. The results indicated that the ViT model achieved the highest accuracy and precision in identifying diseases and assessing ripeness. The models' effectiveness was further enhanced by the integration of attention mechanisms, and the potential for real-world agricultural applications was validated through the use of synthetic data. This research demonstrated that crop monitoring could be significantly improved with advanced machine learning models, particularly ViT, offering a promising tool for more sustainable and efficient strawberry cultivation. Future studies were recommended to expand these methods to other crops and integrate them into broader agricultural practices to enhance productivity and sustainability.