Depth Imaging-Based Framework for Efficient Phenotypic Recognition in Tomato Fruit.

Area of Science:

• Horticultural science
• Computer vision
• Artificial intelligence

Background:

• Tomato is a vital global crop requiring precise phenotyping for breeding and quality control.
• Current phenotyping methods can be labor-intensive and lack high precision.
• Automated, computer vision-based approaches offer potential for efficient and accurate analysis.

Purpose of the Study:

• To develop an intelligent detection framework for automated phenomics analysis of tomato fruits.
• To extract and quantitatively analyze 12 phenotypic traits using image processing and deep learning.
• To provide reliable data for precision tomato breeding and intelligent cultivation.

Main Methods:

• A dataset of tomato fruit section images was created using a depth camera.
• An improved SegFormer model (SegFormer-MLLA) was developed for accurate fruit phenotype segmentation.
• A Hybrid Depth Regression Model fused RGB and depth data for trait estimation.

Main Results:

• The SegFormer-MLLA model achieved accurate segmentation of tomato fruit structures with reduced computational cost.
• The framework accurately detected key phenotypic traits like diameter, thickness, and stem scar dimensions.
• Experimental results showed a high correlation between model-detected and manually measured phenotypic parameters.

Conclusions:

• The developed framework accurately and efficiently automates tomato fruit phenotyping.
• The system provides reliable data supporting precision tomato breeding and intelligent cultivation.
• The methodology serves as a reference for phenotyping other fruit crops.