Integrating deep learning and data fusion for enhanced oranges soluble solids content prediction using machine vision...

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Visible/near-infrared (Vis/NIR) spectral data can be distorted by sample color variations, impacting fruit composition prediction accuracy.
• Accurate prediction of fruit quality attributes like soluble solids content (SSC) is crucial for the food industry.

Purpose of the Study:

• To develop a deep learning model capable of color correction for predicting orange SSC.
• To explore multi-source data fusion of machine vision and Vis/NIR spectroscopy for enhanced prediction accuracy.

Main Methods:

• Designed an integrated device for online acquisition of color images and Vis/NIR transmission spectra.
• Proposed data fusion techniques combining color features and spectral data.
• Constructed color-correction one-dimensional convolutional neural network (1D-CNN) models.

Main Results:

• The optimal color-correction model demonstrated a significant reduction in Root Mean Square Error of Prediction (RMSEP).
• RMSEP decreased by 36.4% compared to partial least squares (PLS) and 16.1% compared to conventional 1D-CNN.
• Multi-source data fusion effectively improved the accuracy of fruit composition prediction.

Conclusions:

• Color correction integrated with deep learning and multi-source data fusion enhances the prediction of fruit soluble solids content.
• Machine vision and Vis/NIR spectroscopy fusion offers a promising approach for accurate and reliable food quality assessment.