DBA-ViNet: an effective deep learning framework for fruit disease detection and classification using explainable AI

Area of Science:

• Computer Vision
• Agricultural Technology
• Plant Pathology

Background:

• Accurate disease identification in fruits is crucial for agricultural productivity and food security.
• Existing computer vision models face challenges in effectively integrating global and local features for precise disease detection.
• The need for robust automated systems in smart agriculture necessitates advanced image analysis techniques.

Purpose of the Study:

• To develop and evaluate a novel computer vision model, the Dual-Branch Attention-Guided Vision Network (DBA-ViNet), for identifying and classifying diseases in multiple fruit types.
• To compare the performance of DBA-ViNet against state-of-the-art pre-trained convolutional neural network (ConvNet) models.
• To enhance the interpretability and trustworthiness of the model's predictions through visualization techniques.

Main Methods:

• Utilized an open-source dataset of fruit disease images (apples, guavas, mangoes, pomegranates, oranges), split into training, validation, and testing sets.
• Implemented 5-fold cross-validation to ensure model generalizability and stability.
• Benchmarked Swin Transformer (ST), EfficientNetV2, ConvNeXt, YOLOv8, and MobileNetV3, and introduced the proposed DBA-ViNet with a dual-branch architecture for integrated feature extraction. Grad-CAM was used for visualization.

Main Results:

• The DBA-ViNet model achieved superior performance, with a testing accuracy of 99.51%, specificity of 99.42%, recall of 99.61%, precision of 99.30%, and F1 score of 99.45%.
• DBA-ViNet outperformed all benchmarked state-of-the-art models across all evaluation metrics.
• Grad-CAM visualizations confirmed that DBA-ViNet accurately focuses on disease-specific symptoms, enhancing model transparency.

Conclusions:

• The proposed DBA-ViNet architecture demonstrates high accuracy and reliability in fruit disease detection.
• Integrating global and local feature extraction via a dual-branch attention mechanism is effective for classification tasks.
• DBA-ViNet shows significant potential for practical application in smart agriculture and automated crop health monitoring systems.