Quantitative Extraction and Evaluation of Tomato Fruit Phenotypes Based on Image Recognition.

Area of Science:

• Agricultural Science
• Computer Vision
• Genetics and Breeding

Background:

• Traditional manual methods for tomato fruit phenotyping are time-consuming and limit high-throughput data collection.
• Accurate measurement of fruit morphology is crucial for tomato breeding and agronomic trait evaluation.

Purpose of the Study:

• To develop automated methods for measuring tomato fruit phenotypes using image recognition and deep learning.
• To enhance the efficiency and precision of data collection for tomato fruit morphology in breeding programs.

Main Methods:

• Selected 10 tomato cultivars with diverse fruit characteristics under controlled illumination.
• Employed image recognition for automated measurement of color and size indicators.
• Utilized a deep learning model (Mask Region-Convolutional Neural Network) for internal structure analysis.

Main Results:

• Successfully extracted key fruit phenotypes including color, dimensions, angles, locule number, and pericarp thickness.
• Achieved an average precision exceeding 0.95 for locule segmentation and counting with the deep learning model.
• Calculated fruit shape index and locule area proportion, demonstrating comparable precision to manual methods with significantly improved efficiency.

Conclusions:

• The combined automated methods provide accurate and efficient tomato fruit phenotyping, overcoming limitations of manual observation.
• This approach offers a valuable tool for high-throughput data collection, reducing artificial errors in breeding.
• The study presents a promising solution for advancing tomato breeding and other fruit crop development through enhanced phenotyping.