DF-SE-ResNet: Orange Leaf Disease Classification Method Based on Improved Attention Mechanism and FPN Fusion

The present study introduces a deep learning-based CNN model, DF-SE-ResNet, designed to enhance the classification precision of orange leaf diseases. This paper employs a feature extraction network that integrates an optimized ResNet50 with FPN (Feature Pyramid Networks) and incorporates an improved attention mechanism module, enhancing the network's ability to extract features from small and dense lesion patches while retaining more semantic information from their surroundings, thus laying a solid foundation for accurate subsequent identification. To address the challenge of capturing lesion patches with random distribution, a region feature weighting fusion method based on permutation invariance is introduced after the feature extraction network, which improves the ability to precisely capture lesion patches and significantly enhances lesion features, effectively resolving the issue of randomly distributed lesion patches. Finally, to mitigate the impact of uneven sample distribution across classes in the dataset, an inverse frequency cross-entropy loss function is adopted in place of the standard cross-entropy loss function, alleviating the model's bias towards the majority classes. In the experimental phase, the DF-SE-ResNet achieved an accuracy rate of 95.4%, demonstrating superior performance compared to other deep learning models.