To handle model bias from long-tailed image datasets, most existing studies have proposed to strengthen the focus on tail classes. However, they usually overlook the intricacies of sample learning hardness and feature norms’ impact on accuracy, as well as the non-target class contribution. We designed a new loss using logit adjustment. It considers the distribution of the number of class samples, sample learning hardness, and feature norms. Moreover, it combines a complementary entropy regularization term designed for long-tailed data. Extensive experiments on long-tailed datasets confirm the proposed loss’s superiority in enhancing classification accuracy.