关键词: 糙米；脂肪酸值；可见图像；偏振图像；可解释性；预测

Abstract: Fatty acid value (FAV) is an important indicator for evaluating the storage quality of paddy rice. Existing techniques for the rapid detection of FAV still struggle to juggle equipment costs, detection throughput, and on-site operational convenience when applied to large-scale grain procurement and storage. Therefore, this study proposed a method for predicting the FAV of brown rice based on the comparative fusion of visible and polarized images and investigated its interpretability. Six representative deep learning architectures, including MobileNetV2, ResNet18, ShuffleNetV2, VGG16, EfficientNetV2, and MobileNetV4, were evaluated for their performance in predicting the FAVs of japonica and indica brown rice samples, and the model with the best overall performance was selected. To further exploit the complementary information contained in visible and polarized images, the discrete wavelet transform was used for pixel-level image fusion, and the FAV prediction performance was evaluated before and after image fusion. Model interpretability was examined by analyzing the associations between image features of brown rice at different FAV levels and kernel hardness, surface color, and multidimensional structural parameters, such as those obtained from micro-computed tomography. The results showed that among the six models, ShuffleNetV2 exhibited the best overall prediction performance for both japonica and indica brown rice. On the validation set, the mean absolute error (MAE) values for the two varieties of brown rice were 1.884 and 1.885 mg/100 g, the root mean square error (RMSE) values were 3.036 and 3.037 mg/100 g, and the coefficients of determination (R2) were 0.712 and 0.861, respectively. The ShuffleNetV2 model based on the fused images also achieved good prediction performance on the test set. For japonica and indica brown rice, the MAE values decreased to 1.378 and 1.670 mg/100 g, the RMSE values decreased to 2.754 and 2.694 mg/100 g, and the R2 values increased to 0.818 and 0.834, respectively. The interpretability analysis showed that for visible-light images, the mean absolute correlation coefficient between FAV and texture features (mean |r| = 0.76) was slightly higher than that between FAV and color features (mean |r| = 0.74), although the difference was marginal. This finding indicated that visible imaging captured textural variations while still effectively characterizing macroscopic changes in kernel color (L*, a*, b*) and overall appearance. In contrast, for polarized images, the mean absolute correlation coefficient between FAV and texture features (mean |r| = 0.87) was substantially higher than that between FAV and color features (mean |r| = 0.65), suggesting that polarization imaging is more sensitive to information related to microstructural changes in rice kernels. In summary, this study provides a novel solution for rapid, non-destructive, and intelligent assessment of rice FAV.

Key words: brown rice; fatty acid value; visible imaging; polarization imaging; interpretability; prediction