Simulation-Driven Spatial Frequency Domain Imaging and Deep Learning for Subsurface Fruit Bruise Discrimination.

Area of Science:

• Agricultural Engineering
• Biophotonics
• Machine Learning

Background:

• Conventional spatial frequency domain imaging (SFDI) optical property inversion is inefficient.
• Deep learning methods for optical property inversion require large-scale real datasets, which are costly and time-consuming to acquire.

Purpose of the Study:

• To develop a simulation-driven approach for subsurface fruit bruise discrimination using SFDI.
• To overcome the limitations of conventional and deep learning methods by generating synthetic datasets.

Main Methods:

• An SFDI simulation environment was built using Blender to generate 800 paired datasets.
• A CBAM-GAN-U-Net model was designed, incorporating surface profile correction to eliminate non-planar distortion.
• The method was validated on liquid phantoms, green apples, and crown pears.

Main Results:

• The CBAM-GAN-U-Net model achieved high accuracy in predicting the reduced scattering coefficient (μs'), outperforming U-Net and GANPOP.
• Non-bruised/bruised fruit discrimination achieved 100% accuracy.
• Mild/severe bruise differentiation showed low misclassification rates (6% for green apples, 8% for crown pears).

Conclusions:

• The simulation-driven approach enables accurate subsurface fruit bruise detection.
• This method provides a reliable technical solution for the fruit and vegetable industry.
• The approach helps reduce postharvest supply chain losses by enabling early detection of fruit damage.