2.5 Loss Function

In our work, the cross-entropy loss function (19) is employed for all deep models. The cross-entropy loss is computed as follows:

where y_i,c is the one-hot encoding format of ground truth labels, p_i,c is the matrix of predicted values for pixels in each class, where the indices, c and i, iterate over all classes and pixels, respectively. Cross-entropy loss is based on minimizing pixel-wise error, where for class imbalanced cases, it may lead to an over-representation for dominant class samples in the loss, resulting in the poor representation or weak contribution for minority samples.