Tobacco leaf disease hole detection method based on spectral–spatial fusion multimodal transformer

Key Points

• This research aims to improve the detection of small lesions in tobacco leaves by integrating CNN and Transformer networks to enhance feature consistency and localization.
• Developed a CNN-Transformer dual-branch network for tobacco defect detection.
• Designed an adaptive fusion mechanism for cross-modal feature consistency.
• Implemented a dynamic detection framework to better localize small lesions.
• Achieved improved localization of small lesions in tobacco leaves with the dual-branch network.
• Demonstrated enhanced feature consistency across modalities, leading to more accurate defect detection.
• Optimized learning for small targets using a multi-scale loss function.