Abstract:

UNet has been widely applied to image segmentation tasks due to its encoder⁃decoder structure and skip⁃connection mechanism. However, traditional downsampling operations (such as max pooling or average pooling) tend to cause loss of spatial detail information during feature compression, while the simple reconstruction approach of the upsampling stage struggles to fully model global contextual semantics. These limitations constrain the network's expressive power and segmentation accuracy. To address this, this paper proposes an enhanced UNet architecture by introducing the Interactive Pooling Module (IPM) and the Global⁃Local Feature Interaction Module (GLFIM) during subsampling and upscaling stages, respectively, to strengthen feature modeling capabilities.Results indicate that in the encoder, the IPM simultaneously integrates max⁃pooling and average⁃pooling branches. Through channel interaction and feature fusion, it effectively enhances feature representation capabilities while preserving more structural and statistical information. In the decoder, the proposed Local Block extracts local details through residual convolutions, while the Global Block combines an attention mechanism based on adaptive pooling with an MLP network to model long⁃range dependencies and global semantic context. Experimental results demonstrate that our method exhibits significant advantages in preserving critical details and enhancing global semantic understanding, achieving superior segmentation performance compared to the standard UNet model. On the WHU dataset, IoU and F1 scores reached 90.73% and 95.14%, respectively.

Key words: Image segmentation, UNet, Interactive convolution, Global-local feature interaction, Feature fusion

摘要：

UNet凭借其编码器⁃解码器结构及跳跃连接机制，被广泛应用于图像分割任务。然而，传统的下采样操作（如最大池化或平均池化）在特征压缩过程中易导致空间细节信息丢失，而上采样阶段的简单重建方式亦难以充分建模全局上下文语义，限制了网络的表达能力和分割精度。为此，提出了一种改进的UNet架构，分别在下采样与上采样阶段引入交互式池化模块（IPM）和全局⁃局部特征交互模块（GLFIM），以增强特征建模能力。结果表明，在编码器部分，IPM同时融合最大池化与平均池化分支，通过通道交互与特征融合，有效提升了特征表示能力并保留了更多结构与统计信息；在解码器部分，提出的局部特征模块通过残差卷积提取细节信息，而全局模块则结合基于自适应池化的注意力机制与多层感知机（MLP）网络，以建模长距离依赖与全局语义上下文；在多个图像分割基准数据集上，本文所提模型在保留关键细节、增强全局语义理解方面具有显著优势，相较于标准UNet模型取得了更优的分割性能；在WHU数据集上，交并比和F1分数分别达到了90.73%和95.14%。

关键词: 图像分割, UNet, 交互式卷积, 全局?局部特征交互, 特征融合