Beyond Single-Scale: A Multi-Scale Approach to Semantic-Enhanced Crop Disease Recognition Through Image–Text Fusion

Crop diseases pose a significant threat to agricultural productivity and global food security. Timely and accurate detection of such diseases is crucial for improving both crop yield and quality. While numerous deep learning approaches rely solely on image data for disease identification, they often overlook the complementary value of multi-scale feature extraction in enhancing visual analysis. To address this limitation and effectively capture hierarchical representations across different scales, we propose a Multi-Scale Feature Extraction Framework for Image-Text Fusion (MSF-ITF) that simultaneously processes both modalities through parallel convolutional branches with varying receptive fields. Our approach utilizes an image encoder to extract foundational visual representations, followed by four parallel convolutional branches (1×1, 3×3, 5×5, 7×7) to capture multi-scale visual information. For text processing, we employ established pipeline using GLM-4V-Plus to generate hierarchical textual descriptions and BLIP2 encoder for feature extraction, then apply multi-scale 1D convolutions to capture semantic patterns at different contextual ranges. Subsequently, a cross-attention module aligns and fuses the multi-scale features from both modalities. Extensive evaluations on Soybean Disease, AI Challenge 2018, and PlantVillage datasets demonstrate that our method outperforms existing single-scale and image-based models in terms of accuracy. Specifically, our model achieves recognition accuracies of 99.14%, 87.90%, and 99.24% on respective datasets, while maintaining computational efficiency through lightweight design. These results highlight the effectiveness of multi-scale learning in integrating hierarchical visual and textual cues for accurate and efficient disease recognition, offering a scalable solution for crop disease diagnosis.