基于DenseNet优化Swin Transformer模型的苹果叶部病害多尺度特征分类

Swin Transformer Model Optimized with DenseNet for Multi-scale Feature Classification of Apple Leaf Diseases

摘要:

针对人工检测苹果病害效率低、成本高且准确性差的问题，本文以Swin Transformer作为基础模型，在核心模块中引入DenseNet思想，增强特征传递和梯度流动；使用Outlook Attention捕捉图像中细节特征，提升模型细粒度信息提取能力。为了进一步优化模型性能，引入了深度可分离卷积和膨胀卷积，实现在较小参数量前提下捕捉不同尺度的特征；在模型中引入Non-Local，以整合全局上下文信息，进一步提高模型的综合性能。以上改进共同作用，使得本文模型在多个任务上表现出了优异的性能和鲁棒性。实验结果显示，苹果叶部病害分类识别准确率达到95.8%，精确率、召回率和F1分数分别达到95.80%、95.74%、95.76%，均超过基线模型。基于改进Swin Transformer的苹果叶部病害分类模型能够有效实现苹果叶部病害的种类识别及其严重程度分类，为大规模作物病害监测提供了理论支持和研究基础，助力精准防控与绿色农业。

Abstract:

Given the limitations of manual apple disease detection, including inefficiency, high cost, and low accuracy, aiming to propose a more effective solution that can improve detection accuracy while reducing time and costs. The Swin Transformer was utilized as the base model and the DenseNet framework was integrated into its core modules to enhance feature propagation and improve gradient flow. Additionally, the Outlook Attention module captured fine-grained image details, enhancing the model’s ability to extract intricate features. To further optimize the model’s performance, the depthwise separable and dilated convolutions were introduced, enabling the capture of multi-scale features while reducing the parameter size. Finally, the Non-Local module was integrated into the model to incorporate global context information, thereby further enhancing overall performance. These improvements collectively enabled the model to exhibit superior performance and robustness across multiple tasks. Experimental results indicated that the accuracy for classifying apple leaf diseases reached 95.8%, with precision, recall, and F1 score values of 95.80%, 95.74%, and 95.76%, respectively, all surpassing those of the baseline model. The proposed Swin Transformer-based model, optimized for apple leaf disease classification, efficiently identified both the type and severity of apple leaf diseases. This served as a theoretical foundation and provided critical support for large-scale crop disease monitoring, facilitating precise disease prevention and control in sustainable agriculture. Moreover, compared with existing deep learning models like ResNet and standard Swin Transformer, the proposed model exhibited superior accuracy and computational efficiency. Future research would focus on further optimizing the model architecture to address more complex agricultural scenarios, such as classifying co-occurring diseases, and integrating drone-based image acquisition technologies for real-time disease detection and prediction.