While deep neural networks are widely used across various domains, particularly in the field of image processing, and have demonstrated impressive performance, they still encounter crucial and unresolved challenges. The problem of catastrophic forgetting is a typical issue faced by most deep learning models at present. This implies that when a model needs to be updated due to changes in its task or to adapt to external conditions by acquiring new training data for new categories, it necessitates retraining for the new phase. However, the old training data is no longer accessible. In recent years, a series of solutions and incremental learning frameworks have been proposed in the academic field to address the issue of catastrophic forgetting caused by the need for continuous updates to current models. However, these solutions largely focus on classification and detection tasks and have shown some effectiveness. In this paper, we aim to tackle this issue within the context of semantic segmentation, which involves addressing pixel-level classification problems. We concern the issue of background shift caused by the unavailability of labels for old classes in the new training phase, and have utilized a modified cross-entropy loss function. Simultaneously, we employ dual distillation on intermediate features and outputs to capture the knowledge learned by the old model, thereby constraining the training process of the new model.