Fig. 3

Residual network structure. The residual network structure uses cross-layer connections to directly pass input signals to subsequent layers and modify them in later layers to learn residual information. This network structure makes the training of SAFNet easier, while also improving its performance and convergence speed. In addition, residual network structure can effectively reduce the number of model parameters and improve model generalization ability. (SAFNet denotes scene aware fusion network; DSC denotes Dice Similarity Coefficient.)