A Image Segmentation Network Based on Dual Parallel Encoders

With the advancements in deep learning over recent years, deep learning-driven image processing has achieved remarkable breakthroughs in tasks such as semantic segmentation, object detection, and generative modeling. CNN-based image segmentation algorithms are difficult to capture remote dependencies due to their higher focus on local features. Transformer-based methods are more capable of modeling contextual dependencies due to their greater focus on global features, yet lack local details. Thus inspired by this, this paper first proposes Double Encoder UNet (DEUNet), which is a U-like network that utilizes CNN and Transformer to build a dual parallel encoder. To effectively integrate global and local features, we first propose a novel feature fusion module that enables organic combination of multi-dimensional features. Additionally, an attention mechanism integrated with skip connections is introduced to enhance contextual modeling of image semantics. Finally, experiments are performed on widely used public datasets (DRIVE and ISIC), where DEUNet achieves superior segmentation performance. The results highlight the advantages of parallel coding via CNN-Transformer architectures in capturing hierarchical visual patterns.