关键词: 柑橘, 格拉姆角和场, 深度学习, 可见—近红外光谱, 卷积注意力机制

Abstract: Aiming at the problems of spectral information loss and low detection accuracy in the current citrus fruit internal quality inspection model, a new method based on Gramian Angular Summation Field (GASF) transformation and deep learning was proposed for non‑destructive qualitative analysis of internal citrus fruit quality. Through GASF transformation, the collected one‑dimensional spectra data in the visible‑near infrared range of citrus fruits are converted into two‑dimensional images. The data are augmented by  Moving Average (MA), Standard Normal Variate (SNV) and others preprocessing methods. A two‑dimensional Convolutional Neural Network (2D—CNN) model is designed, incorporating a Convolutional Block Attention Module (CBAM) to enhance the model's feature extraction capabilities from GASF images. Experimental results reveal that, compared to traditional machine learning models such as Support Vector Machine (SVM) and Random Forest (RF), the neural network model exhibits stronger spectral information extraction capability and higher predictive accuracy. The prediction accuracy of SVM and RF are 84.85% and 81.82% respectively. The predictive accuracy of 2D—CNN is reported as 87.88%, which increases to 93.94% with the incorporation of CBAM. The study indicates that GASF transformation effectively introduces the advantages of neural networks in image processing into visible‑near infrared spectral analysis, offering new insights and theoretical references for non‑destructive internal fruit quality inspection.

Key words: citrus, Gramian Angular Summation Field, deep learning, visible?near infrared spectrum, convolution attention mechanism