Scholarlink

<p>Prompt and reliable plant disease identification is a&#xa0;highly significant aspect of precision agriculture and environmentally friendly crop management practices. Unfortunately, many visual deep learning-based methods have been solely dependent on appearances, thereby limiting their capabilities to capture subtle semantic distinctions among visually similar disease categories. This paper introduces a&#xa0;lightweight cross-modal apple disease recognition (CM-ADR) approach that co-integrates visual and textual representations for the fine-grained classification of apple leaf diseases. To be more specific, the proposed system architecture uses a&#xa0;Swin Transformer-based visual encoder together with a&#xa0;Bootstrapping Language–Image Pre-training&#xa0;2 (BLIP-2) text encoder to capture hierarchical lesion-aware visual features and semantic textual embeddings that describe apple disease characteristics, morphology, and texture patterns. Besides, a&#xa0;multi-layer cross-attention fusion mechanism is proposed to establish semantic alignment between visual and textual modalities, thereby leading to robust cross-modal feature learning and enhanced disease discrimination. The proposed structure’s effectiveness is demonstrated through experiments using four different public datasets: FGVC7, PlantVillage Apple, Kaggle Apple Leaves, and Apple Tree Leaf Disease Segmentation Dataset (ATLDSD). The results show that CM-ADR achieves excellent classification accuracies of 99.78%, 87.6%, 91.2%, and 88.5% for each dataset, respectively, while having only 1.14 million trainable parameters, which is much less than the number of trainable parameters in regular convolutional neural network and Transformer-based models, while also surpassing them in both accuracy and computational efficiency. In addition, Gradient-weighted Class Activation Mapping visualization and quantitative interpretability analysis show that the suggested model is capable of generating highly focused and semantically meaningful attention maps, with the highest localization accuracy reaching 92.3% and the activation focused on lesions being better than that of existing lightweight architectures. These results also prove that semantically guided textual descriptions greatly help fine-grained disease identification and improve model explainability. Our proposed system is an innovative, scalable, explainable, and edge-artificial intelligence-friendly method for smart plant disease diagnosis and future precision agriculture.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Cross-Modal Apple Disease Recognition: Enhancing Apple Leaf Disease Detection Through Vision–Language Feature Alignment

  • 期刊论文

  • 发表日期：

    2026年7月13日

查看全文
  • Vinay Gautam,
  • Jyoti Rani

摘要

Prompt and reliable plant disease identification is a highly significant aspect of precision agriculture and environmentally friendly crop management practices. Unfortunately, many visual deep learning-based methods have been solely dependent on appearances, thereby limiting their capabilities to capture subtle semantic distinctions among visually similar disease categories. This paper introduces a lightweight cross-modal apple disease recognition (CM-ADR) approach that co-integrates visual and textual representations for the fine-grained classification of apple leaf diseases. To be more specific, the proposed system architecture uses a Swin Transformer-based visual encoder together with a Bootstrapping Language–Image Pre-training 2 (BLIP-2) text encoder to capture hierarchical lesion-aware visual features and semantic textual embeddings that describe apple disease characteristics, morphology, and texture patterns. Besides, a multi-layer cross-attention fusion mechanism is proposed to establish semantic alignment between visual and textual modalities, thereby leading to robust cross-modal feature learning and enhanced disease discrimination. The proposed structure’s effectiveness is demonstrated through experiments using four different public datasets: FGVC7, PlantVillage Apple, Kaggle Apple Leaves, and Apple Tree Leaf Disease Segmentation Dataset (ATLDSD). The results show that CM-ADR achieves excellent classification accuracies of 99.78%, 87.6%, 91.2%, and 88.5% for each dataset, respectively, while having only 1.14 million trainable parameters, which is much less than the number of trainable parameters in regular convolutional neural network and Transformer-based models, while also surpassing them in both accuracy and computational efficiency. In addition, Gradient-weighted Class Activation Mapping visualization and quantitative interpretability analysis show that the suggested model is capable of generating highly focused and semantically meaningful attention maps, with the highest localization accuracy reaching 92.3% and the activation focused on lesions being better than that of existing lightweight architectures. These results also prove that semantically guided textual descriptions greatly help fine-grained disease identification and improve model explainability. Our proposed system is an innovative, scalable, explainable, and edge-artificial intelligence-friendly method for smart plant disease diagnosis and future precision agriculture.