To address the issues of low accuracy and practicality in existing tobacco leaf grading models, which are mostly built upon the front-side features of flattened tobacco leaves, a multi-site front and back recognition method was proposed using a genetic algorithm-regularized extreme learning machine(GA-RELM). Firstly, multi-scale features from both side of tobacco leaves in their natural state were extracted to construct dataset. Feature importance and relationships were analyzed to reduce dimensionality and construct optimized feature combinations. Secondly, the hidden layer biases of the regularized extreme learning machine(RELM) were optimized to enhance model accuracy and applicability. The results showed that compared with the original extreme learning machine(ELM), GA-RELM improved classification accuracy by 0.84% and 7.88% for front/back leaves and other parts, respectively, while reducing the computation time by 2.56 s and 5.72 s, respectively. GA-RELM outperformed other grading algorithms in accuracy, precision, recall and F1-score.