The long-tailed distribution data poses many challenges for machine learning because the tail classes are extremely scarce. Long-tailed data augmentation is a powerful technique for enriching the tail class diversity. However, existing methods often treat each class independently, assuming that classes are isolated from each other. These approaches overlook the presence of easily confused tail classes, making it challenging for models to distinguish between them accurately. In this paper, we propose a long-tailed classification method based on data augmentation, which utilizes multi-granularity knowledge to select and combine easily confused tail samples, thereby enhancing the classification performance of these samples. First, we utilize multi-granularity knowledge and semantic relation trees to build a class relation matrix. This matrix records the relationship between classes and helps the model search for easily confused classes from bilateral branch samplers. Second, we crop and combine the easily confused head and tail class samples in a foreground–background manner to generate new samples, thereby augmenting the model training. The extensive head class knowledge is transferred to the scarce tail class samples through the combination of fore-background, and the discriminative and generalized abilities of the model are improved. The experimental results affirm the effectiveness of our proposed method.