Plant Root Phenotyping Using Deep Conditional GANs and Binary Semantic Segmentation.

Area of Science:

• Plant Biology
• Computer Vision
• Bioinformatics

Background:

• Accurate plant root phenotyping is crucial for crop improvement and understanding plant responses.
• Pixel-wise class imbalance in root images poses a significant challenge for semantic segmentation algorithms.
• Existing methods struggle with generating high-resolution, realistic root images and annotations.

Purpose of the Study:

• To develop an effective binary semantic segmentation approach for Arabidopsis thaliana root images.
• To address pixel-wise class imbalance using conditional generative adversarial networks (cGANs).
• To enhance plant root phenotyping through improved image segmentation.

Main Methods:

• Utilized Pix2PixHD, an image-to-image translation cGAN, to generate realistic, high-resolution root images and annotations.
• Augmented the original root dataset by tripling its size using the trained cGAN to mitigate class imbalance.
• Employed SegNet for semantic segmentation of root pixels, followed by postprocessing to refine segmentation results.

Main Results:

• The cGAN successfully generated high-resolution, realistic root images and annotations.
• Dataset augmentation significantly reduced pixel-wise class imbalance.
• The SegNet model achieved high testing accuracy (>99%), low cross-entropy error (<2%), and a high Dice Score (≈0.80).
• The segmentation approach demonstrated low inference time, enabling near real-time processing.

Conclusions:

• Conditional generative adversarial networks (cGANs) are effective for generating high-quality plant root images and annotations.
• The proposed method successfully reduces pixel-wise class imbalance, leading to improved semantic segmentation accuracy.
• This approach offers a robust and efficient solution for plant root phenotyping, achieving state-of-the-art performance.