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

Abstract

Conventional spatial frequency domain imaging (SFDI) based optical property inversion is inefficient, while deep learning methods suffer from heavy reliance on large-scale real datasets. To address this contradiction, a simulation-driven approach for subsurface fruit bruise discrimination was proposed. An SFDI simulation environment was built with Blender to generate 800 paired datasets of diffuse reflectance images and optical transport coefficients, overcoming the high cost and long cycle of real dataset acquisition. We designed the CBAM-GAN-U-Net model and adopted surface profile correction in the prediction method to eliminate curved surface-induced non-planar distortion, with the whole method validated on liquid phantoms, green apples and crown pears. This prediction method achieved high accuracy in predicting the reduced scattering coefficient μ s ', with NMAE of 0.021 ± 0.007 (phantoms), 0.039 ± 0.012 (severely bruised green apples) and 0.044 ± 0.015 (severely bruised crown pears), outperforming U-Net and GANPOP. Based on the predicted μ s ', a discrimination strategy combining coefficient of variation, mean ratio and receiver operating characteristic (ROC) curve analysis was adopted, attaining 100% accuracy for non-bruised/bruised fruit discrimination, with misclassification rates of 6% (green apples) and 8% (crown pears) for mild/severe bruise differentiation. This method enables accurate subsurface fruit bruise detection, providing a reliable technical solution for the fruit and vegetable industry and helping reduce postharvest supply chain losses.

Plant phenotyping relevance

SFDIと深層学習による果実内部の打撲状態・重症度の画像推定手法を開発し、ファントムと果実で検証しており、植物（果実）の状態取得が中心である。

Code and data availability

The supplied article blocks describe simulated SFDI datasets (Blender), a CBAM-GAN-U-Net model, and real fruit/phantom measurements, but contain no data or code availability statement, no public repository, and no author-provided URL for datasets, images, code, or trained models. No qualifying paper-specific public or

No evidence-backed public reproduction asset is currently recorded.