Related Experiment Video
Updated: Jun 23, 2026
12:08
From Voxels to Knowledge: A Practical Guide to the Segmentation of Complex Electron Microscopy 3D-Data
Published on: August 13, 2014
LViM: Language-Infused Visual Mamba for apple leaf pests and diseases precise segmentation in complex environments
Jiale Chen1, Wei Shi1, Ziyang Shi1
1School of Electronic Information and Physics, Central South University of Forestry and Technology, Changsha, 410004, China.
Plant Phenomics (Washington, D.C.)
|June 22, 2026
Summary
A new deep learning model, Language-Infused Visual Mamba (LViM), enhances apple leaf disease segmentation by integrating text and visual data. This approach improves accuracy in complex orchard conditions, crucial for preserving crop yield and quality.
Area of Science:
- Agricultural Science
- Computer Vision
- Deep Learning
Background:
- Accurate apple leaf disease segmentation is vital for maintaining crop yield and quality in global fruit production.
- Real-world orchards present challenges including low contrast lesions, leaf overlap/occlusion, and limitations of RGB imagery for subtle feature detection.
- Existing deep learning models struggle with the complexities of natural orchard environments, impacting segmentation accuracy and generalization.
Purpose of the Study:
- To develop an advanced deep learning model for robust apple leaf disease segmentation in challenging orchard conditions.
- To address limitations in lesion localization, feature representation, and pathological feature identification using unimodal imagery.
- To improve the accuracy and generalization capabilities of automated plant disease detection systems.
Main Methods:
- Proposed Language-Infused Visual Mamba (LViM), a dual-path U-Net architecture integrating Mamba and Transformer modules for semantic-visual fusion.
- Developed a U-shaped Multimodal Transformer (MTT) branch with AMBERT for enhanced textual feature extraction and semantic cue integration.
- Implemented a U-shaped Visual State Space (VMamba) branch using Selective Scanning (SS2D) and Visual State Space (VSS) blocks for global context and fine-grained detail capture.
- Utilized Cross-Attention Gate Fusion (CAGF) and Linguistic Cross-Nested (LCN) modules for efficient cross-modal alignment and hierarchical feature modeling.
Main Results:
- LViM demonstrated superior performance compared to the VM-UNet baseline in complex orchard environments.
- Achieved significant improvements: 4.05% in Precision, 4.25% in Dice coefficient, 4.49% in mean Intersection over Union (mIoU), and 4.23% in Recall.
- The model effectively mitigated issues related to low contrast, leaf occlusion, and subtle pathological feature detection.
Conclusions:
- The proposed LViM model offers a robust solution for apple leaf disease segmentation, outperforming existing methods.
- Integrating language and visual information through the LViM architecture enhances segmentation accuracy and model generalization.
- This approach holds significant potential for improving automated disease monitoring and management in apple orchards.
