A deep learning method with cross dropout focal loss function for imbalanced semantic segmentation

Deep learning methods have proven their potential in semantic segmentation. However, they depend on the data quality and training process. Often, the data corresponding to the objects to be segmented are of different sizes and this creates difficulties for the segmentation method. Objects are segmented and associated with categories during the training process. Data imbalance is a challenging problem, which often results in unsatisfactory segmentation performance. This paper proposes a solution to this task based on a novel cross dropout focal loss (CDFL) function, which represents well the change between the cross-entropy and other state-of-the-art loss functions providing a balance between the precision and accuracy of segmentation. The performance of the considered fully convolutional network (FCN) with different loss functions is considered and carefully evaluated. The proposed loss function improves efficiently the semantic segmentation performance over other well-known loss functions. It is demonstrated on Cityscapes and PASCAL VOC 2010 publicly available datasets. The implementation is over relatively large data sets. The achieved mean accuracy of the proposed CDFL network on Cityscapes dataset is 76.41% and on PASCAL VOC 2010 dataset is 79.63% which is with approximately 2.5% improvement compared with the same network implemented with the cross-entropy loss function.