“…For imbalanced datasets, in addition to the RES, another strategy is to assign different weights according to the number of various classes to improve model performance. Sivapuram et al [35] designed a loss function called visually interpretable space adjustment learning (VISAL), which can create more space for generalization of minority class samples by introducing angular and Euclidean margins in the cross-entropy learning strategy. Tian et al [36] proposed a hard class mining loss, which reshapes the CEL by dynamically weighting the loss of each class based on instantaneous recall performance.…”