The soluble solids content (SSC) is an important index for evaluating the quality of oranges. However, because of the complex internal organizational structure of oranges, different tissues may have a significant impact on the incident light, which makes it difficult to construct a high-precision and stable model for SSC prediction. In this study, full-transmittance hyperspectral imaging technology was used to collect information on the entire orange. The raw Vis-NIR hyperspectral data were encoded into GAF images and the image features were extracted using HOG operators. Finally, the optimised GAF-HOG-SVR model obtained satisfactory prediction accuracy, with a correlation coefficient of 0.927 and a root mean square error of 0.445 for the prediction set. This study demonstrates that the proposed method can effectively overcome the adverse effects of complex internal tissues in oranges on SSC prediction, providing a new approach for the accurate and stable nondestructive quality evaluation of oranges.