Cotton leaf segmentation with composite backbone architecture combining convolution and attention.

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Accurate plant leaf segmentation and edge recognition are crucial for automated plant phenotyping.
• Current semantic segmentation models require significant computational resources and expert adjustments for specific applications like cotton leaf analysis.

Purpose of the Study:

• To develop a simple, effective semantic segmentation architecture for cotton leaf segmentation.
• To reduce the trial-and-error costs associated with adapting segmentation models for agricultural use.

Main Methods:

• Proposed a novel semantic segmentation architecture featuring a composite backbone.
• The composite backbone integrates CoAtNet (combining Transformer attention with convolution) and Xception.
• The model leverages both convolutional neural network (CNN) operations and Transformer attention mechanisms.

Main Results:

• The proposed model outperformed benchmark models (PSPNet, DANet, CPNet, DeepLab v3+) on cotton leaf segmentation.
• Achieved high performance in leaf edge segmentation with MIoU of 0.940 and BIoU of 0.608.
• The composite backbone effectively balances performance with reduced computational demands compared to pure Transformer models.

Conclusions:

• The developed model offers an efficient and effective solution for cotton leaf segmentation and edge recognition.
• It alleviates the computational burden of Transformer-based models while maintaining high accuracy.
• Provides a viable scheme for high-throughput plant phenotypic feature detection in agriculture.