Research on Semantic Segmentation Method for Maritime Navigation Scene Based on Improved DMNet

In recent years, semantic segmentation has been continuously developing, but there are still problems such as incomplete image understanding, incomplete contextual information extraction, imbalanced data samples, and uneven object distribution. A semantic segmentation algorithm based on an improved Dynamic Multi scale Network (DMNet) is proposed to address the current issues. An attention mechanism is added to the decoder to make the network pay more attention to the target to be detected; Using online difficult sample mining to improve the problem of sample imbalance in the dataset; Utilize the CrossEntropyLoss+Lovasz SoftmaxLoss joint loss function and the Adam with decoupled weight Decay (AdamW) optimizer to improve network training effectiveness and segmentation accuracy. The algorithm achieves 85.01% mIoU and 90.30% mAcc, effectively optimizing the model structure and improving segmentation accuracy, thus providing a certain technical foundation and application reference for semantic understanding of maritime channel scenes.