全文下载:
202602017.pdf
文章编号:1672-6987(2026)02-0114-07;DOI:10. 16351/j. 1672-6987. 2026. 02. 017
王程远,高 林* (青岛科技大学 自动化与电子工程学院,山东 青岛 266061)
摘 要:针对传统的 UNet 网络对皮肤病变区域的错误识别、分割精度低以及复杂区域难以分 割等问题,提出一种融合算法。该算法利用门控完全特征融合模块和空间注意力模块及 Trans⁃ former 的改进来解决上述问题。首先在 UNet编码器和解码器中集成自适应卷积块,用以捕捉 病变区域中的几何变换。然后在编码解码连接时使用门控完全特征融合模块用以衡量各个特 征图的重要内容。在解码器之前设计了密集空间注意力模块,生成更具判别性的特征内容。最 后使用 Transformer 特征提取系统,提高特征在空间中的映射能力。与传统的 UNet网络相比, 提高了系统 Dice系数、准确率、敏感度、交并比指标,是一种有效的基础网络结构。
关键词:自适应卷积块;门控完全特征融合模块;空间注意力模块;改进 UNet模型
中图分类号:TP 751. 1 文献标志码:A
引用格式:王程远,高林 . 基于改进 UNet 模型的皮肤病分割算法[J]. 青岛科技大学学报(自 然科学版),2026,47(2):114-120.
WANG Chengyuan, GAO Lin. Dermatological segmentation algorithm based on improved UNet model[J]. Journal of Qingdao University of Science and Technology(Natural Science Edition),2026,47(2):114-120.
Dermatological Segmentation Algorithm Based on Improved UNet Model
WANG Chengyuan,GAO Lin (College of Automation and Electronic Engineering, Qingdao University of Science and Technology, Qingdao 266061,China)
Abstract:In this paper, a fusion technique is proposed for the traditional UNet network to mis⁃ identify skin lesions, low segmentation accuracy, and difficult segmentation of complex areas. This technology uses improvements in the gated full feature fusion module and spatial attention module and Transformer to solve the above problems. First, modulated deformable convolutional blocks are integrated in UNet encoders and decoders to capture geometric transformations in the lesion area. Then, the gated full feature fusion module is used to measure the important content of each feature map when encoding and decoding the connection, and the dense spatial attention module is designed before the decoder to generate more discriminative feature content. Finally, the Transformer feature extraction block system is used to improve the mapping ability of features in space. Compared with the traditional UNet network, improving the Dice coefficient, accuracy, sensitivity, and cross-union ratio of the system is an effective basic network structure.
Key words: adaptive convolution module; gated fully feature module; spatial attention module; improved UNet model
收稿日期:2025-05-15
基金项目:国家自然科学基金项目(61971253).
作者简介:王程远(1998—),男,硕士研究生 . * 通信联系人 .