Semi-Self-Supervised Learning for Semantic Segmentation in Images with Dense Patterns.

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Manual image annotation for deep learning is costly and time-consuming, especially for complex tasks like semantic segmentation of plant images.
• Pixel-level annotations for dense, irregularly patterned objects, such as wheat heads, present significant challenges.

Purpose of the Study:

• To develop a high-performing deep learning model for semantic segmentation with minimal manual annotation.
• To address the cost and time barriers associated with large-scale annotated datasets in agricultural imaging.

Main Methods:

• Synthesized a computationally annotated dataset using a few manually annotated images and unannotated video clips.
• Employed a customized U-Net model trained on synthesized data.
• Applied three domain adaptation steps to bridge the distribution gap between synthesized and real images.
• Fine-tuned the model with additional annotated images from diverse domains.

Main Results:

• Achieved a Dice score of 0.89 on an internal test set using only two annotated images.
• Obtained a Dice score of 0.73 on a diverse external dataset from 18 domains.
• Improved the Dice score to 0.91 after fine-tuning with domain-specific data.

Conclusions:

• The proposed method enables high-performance semantic segmentation with significantly reduced manual annotation effort.
• The approach is effective for tasks involving irregularly repeating object instances, such as wheat head segmentation.
• The methodology is adaptable to other segmentation challenges where large annotated datasets are scarce.