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

The primary aim of this research is to develop an effective and robust model for identifying and classifying diseases in general fruits, particularly apples, guavas, mangoes, pomegranates, and oranges, utilizing computer vision techniques.

An open-source collection of fruit disease images, comprising both diseased and healthy samples from the first five fruit types, was used in this study. The data was split into 70% training, 15% validation, and 15% testing. A 5-fold cross-validation was used to maintain the generalizability and stability of the model's performance.

For performance comparisons of these models on the dataset, we benchmarked state-of-the-art pre-trained convolutional neural network (ConvNet) models, including Swin Transformer (ST), EfficientNetV2, ConvNeXt, YOLOv8, and MobileNetV3. A new model, the Dual-Branch Attention-Guided Vision Network (DBA-ViNet), was introduced. A hybrid with two branches of DBA-ViNet can efficiently integrate global and local features for improved disease identification accuracy. Grad-CAM was used to visualize the regions that contributed to each prediction, helping to interpret the model. These heatmaps verified that DBA-ViNet can correctly direct its attention to disease-specific symptoms, thereby increasing trust and transparency in the classification results.

The proposed DBA-ViNet achieved a high testing classification accuracy of 99.51%, specificity of 99.42%, recall of 99.61%, precision of 99.30% and F1 score of 99.45% outperforming baseline models in all evaluation metrics. While the improvements were consistent, statistical significance testing was not performed and will be explored in future work.

These results confirm the effectiveness of the proposed DBA-ViNet architecture in fruit disease detection, suggesting that incorporating both global and local feature extraction into the design of the double-branch attention mechanism for classification can achieve high accuracy and reliability. It is potentially practical in smart agriculture and the automated crop health monitoring system.