Related Experiment Videos
DeepLabV3+ with MobileNetV3 backbone enhancement and multi-level fusion decoder for pepper field weed segmentation
Jun Lu1,2, Fan Wang1, Donglin Cao2
1School of Intelligent Manufacturing and Robotics, Shanghai Dianji University, Shanghai, China.
Frontiers in Plant Science
|August 15, 2026
Summary
This study introduces an advanced semantic segmentation model for precise crop-weed identification in pepper fields. The improved framework enhances accuracy and efficiency for practical precision weeding applications.
Area of Science:
- Agricultural Engineering
- Computer Vision
- Machine Learning
Background:
- Accurate crop-weed segmentation is crucial for effective precision weed management in pepper cultivation.
- Challenges include similar plant textures, occlusion, varying illumination, and small, dispersed weed patches.
Purpose of the Study:
- To develop an improved semantic segmentation framework for accurate pepper-weed discrimination.
- To enhance the model's ability to handle complex field conditions and improve boundary detection.
Main Methods:
- An enhanced DeepLabV3+ framework utilizing MobileNetV3 backbone.
- Integration of Adaptive Activation Fusion, Efficient Channel Attention, and a customized atrous spatial pyramid pooling module.
- Application of focal cross-entropy and Dice loss with weighted sampling, class-aware cropping, and data augmentation.
Main Results:
- Achieved a mean Intersection over Union (mIoU) of 95.87% and recall of 97.98%.
- Obtained a boundary F1-score of 95.88% and boundary IoU of 92.13%.
- Demonstrated an efficient inference speed of 46.86 FPS.
Conclusions:
- The proposed framework significantly improves region-level segmentation accuracy and boundary delineation.
- The model offers enhanced inference efficiency, showing strong potential for real-world precision weeding in pepper fields.
