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

Abstract

Tomato is a globally significant horticultural crop with substantial economic and nutritional value. High-precision phenotypic analysis of tomato fruit characteristics, enabled by computer vision and image-based phenotyping technologies, is essential for varietal selection and automated quality evaluation. An intelligent detection framework for phenomics analysis of tomato fruits was developed in this study, which combines image processing techniques with deep learning algorithms to automate the extraction and quantitative analysis of 12 phenotypic traits, including fruit morphology, structure, color and so on. First, a dataset of tomato fruit section images was developed using a depth camera. Second, the SegFormer model was improved by incorporating the MLLA linear attention mechanism, and a lightweight SegFormer-MLLA model for tomato fruit phenotype segmentation was proposed. Accurate segmentation of tomato fruit stem scars and locular structures was achieved, with significantly reduced computational cost by the proposed model. Finally, a Hybrid Depth Regression Model was designed to optimize the estimation of optimal depth. By fusing RGB and depth information, the framework enabled efficient detection of key phenotypic traits, including fruit longitudinal diameter, transverse diameter, mesocarp thickness, and depth and width of stem scar. Experimental results demonstrated a high correlation between the phenotypic parameters detected by the proposed model and the manually measured values, effectively validating the accuracy and feasibility of the model. Hence, we developed an equipment automatically phenotyping tomato fruits and the corresponding software system, providing reliable data support for precision tomato breeding and intelligent cultivation, as well as a reference methodology for phenotyping other fruit crops.

Code and data availability

The paper's Data Availability Statement explicitly deposits some datasets, model weights, and code used in the study at a public GitHub repository, which is paper-specific and actionable. Full self-developed datasets require contacting the corresponding author.