Enhancing Semantic Segmentation Through a Targeted Approach to Hard-to-Classify Examples

This research paper examines the effectiveness of semantic segmentation utilizing a weighted cross-entropy loss function. A modified version of cross-entropy loss is implemented to address data imbalance and improve predictions for hard-to-classify classes. Hyperparameters of the deep learning model are adjusted based on this modified loss function. The performance of the proposed loss function is evaluated through standard metrics.