Related Experiment Video
Updated: Mar 20, 2026
11:38
Author Spotlight: Enhancing PSC-to-Functional Cell Differentiation Using ML Models Based on Live-Cell Bright-Field Imaging
Published on: October 4, 2024
1.2K
CMNet: an asymmetric dual-branch network for accurate cotton segmentation.
Gengrong Zhang1, Halidanmu Abudukelimu1, Mayilamu Musideke1
1School of Information Management, Xinjiang University of Finance and Economics, Urumqi, Xinjiang, China.
Frontiers in Plant Science
|March 19, 2026
Summary
A new deep learning model, CMNet, improves cotton segmentation accuracy in complex agricultural fields. This advanced network enhances precision for tasks like intelligent harvesting and yield estimation, outperforming existing methods.
Area of Science:
- Agricultural Automation
- Computer Vision
- Deep Learning
Background:
- Precise cotton segmentation is crucial for agricultural automation tasks like harvesting and yield estimation.
- Complex field environments with background interference and irregular shapes challenge existing deep learning segmentation methods, leading to inaccuracies.
- Current methods often suffer from insufficient accuracy, over-segmentation, and misidentification in cotton segmentation.
Purpose of the Study:
- To develop a novel and efficient deep learning model for accurate cotton segmentation in challenging agricultural settings.
- To address limitations of existing methods, including accuracy, over-segmentation, and misidentification.
- To improve the extraction of local details and global semantic information while reducing computational load.
Main Methods:
- Proposed Cotton-aware Mamba-enhanced UNet (CMNet), optimizing ParaTransCNN architecture.
- Integrated 2D Selective Scan (SS2D) module to balance local/global information extraction and reduce computation.
- Incorporated Deformable Convolutional Networks v1 (DCNv1) into the Vision Mamba (VMamba) branch for enhanced boundary delineation.
- Added Atrous Spatial Pyramid Pooling (ASPP) to the Convolutional Neural Network (CNN) branch for multi-scale feature representation.
- Utilized Spatial and Channel Squeeze-and-Excitation (scSE) attention mechanism for improved feature modeling.
Main Results:
- CMNet achieved superior performance on an in-field cotton image dataset, with Dice (91.06%), mIoU (84.18%), and Accuracy (98.10%).
- The model demonstrated reduced parameter count and computational complexity compared to existing methods.
- Generalization experiments on other plant datasets showed outstanding results, confirming CMNet's adaptability.
Conclusions:
- CMNet offers a significant advancement in cotton segmentation accuracy and efficiency for agricultural automation.
- The model's adaptability suggests potential for broader applications in multi-crop segmentation.
- This research provides valuable insights for smart agriculture segmentation and contributes to the field with publicly available code and data.
