Computer Vision and Field Validation of an Artificial Intelligence-Based Tomato Grader for Large-Scale Production in Northeastern Thailand

Abstract

Tomato (Solanum lycopersicum L.) quality grading based on visual inspection often yields inconsistent results and reduces market value. The 'Perfect Gold 111' variety presents distinct morphological traits, including a characteristic green-to-red color transition, specific calyx structure, and defect patterns such as greenback distribution and suberization. These characteristics differ substantially from internationally studied cultivars, rendering generic pre-trained models insufficient for accurate grading under the Thai TACFS 1503–2007 standard. This study developed an automated grading model for 'Perfect Gold 111' tomatoes using a deep learning model based on a flow-based (node-based) architecture integrated with the Robot Operating System (ROS) framework and implemented on the CiRA CORE platform. A total of 220 samples were collected and graded according to the TACFS 1503–2007 standard. Top and side-view images were used to create a dataset comprising 165 tomatoes for training and 55 for testing. Model performance was evaluated using Precision, Recall, F1-Score, and Accuracy, and was compared with manual grading performed by farmers. The AI model achieved 80.00% of accuracy, outperforming farmer grading, which achieved 52.72% accuracy.  In addition, the model reduced misclassification among visually similar grades and provided consistent, quantitative assessments of color, shape, and defects. These findings highlight the potential of AI-based grading systems to improve quality consistency, reduce labor, and support automated postharvest sorting for both smallholder and industrial tomato production.