SweetFruit: A Two-Stage Mobile Sensing System for Real-Time Fruit Sugar Estimation

2026-05-31

AI Summary

This study addresses the limitations of conventional fruit sugar content (Brix) assessment methods, which are often destructive, time-consuming, or require physical contact, thereby hindering high-throughput agricultural quality inspection. To overcome these challenges, the authors propose a low-cost, two-stage, non-contact mobile system. In the first stage, a time-of-flight (ToF) depth camera coupled with a lightweight SF-PointNet model rapidly classifies fruits into high- or low-sugar categories using 3D point clouds. The second stage integrates 18-channel near-infrared spectral data with geometric information from the initial screening in a task-oriented multimodal framework to enhance regression accuracy, employing a compact SF-Net for precise Brix prediction. Implemented on an embedded platform for real-time operation, the system achieves over 90% classification accuracy and a Brix prediction RMSE of 0.57 on green apples and strawberries, reducing error by 22% compared to spectral-only approaches.

Abstract

Accurate prediction of fruit sugar content is essential for quality control and market valuation in agriculture. Conventional measurement techniques rely on destructive, time-consuming processes (e.g., juicing and refractometry) or direct contact instruments, which hinder high-throughput operations. This paper introduces SweetFruit, a mobile two-stage system that leverages low-cost sensors to estimate fruit sugar content without contact. In Stage 1, we implement a lightweight 3D deep learning model (SF-PointNet) that uses point clouds from a Time-of-Flight (ToF) depth camera to classify fruit as high or low sugar. In Stage 2, a regression network (SF-Net) predicts the fruit's Brix value using measurements from a compact 18-channel near-infrared (NIR) spectrometer. The system uses simple off-the-shelf sensors (AS7265x NIR and Arducam ToF) with efficient processing pipelines for real-time execution on embedded platforms. Experiments on green 'Granny Smith' apples and strawberries demonstrate the system's effectiveness. Stage 1 achieves over 90% classification accuracy, enabling rapid prescreening, while Stage 2 delivers precise sugar estimates, with a root mean square error (RMSE) of 0.57 Brix, reducing error by 22% compared to using NIR sensing alone. SweetFruit offers a scalable, field-ready solution for rapid fruit quality screening, showcasing the benefits of task-specific multimodal sensing in mobile agricultural applications.

Problem

• fruit sugar estimation
• non-destructive sensing
• real-time measurement
• quality control
• agricultural application

Innovation

• multimodal sensing
• two-stage deep learning
• non-destructive sugar estimation
• mobile agricultural system
• real-time Brix prediction