非對(duì)稱的編碼器-解碼器架構(gòu)下圖像分割方法研究

打開(kāi)文本圖片集

中圖分類號(hào)：TP391.41 文獻(xiàn)標(biāo)志碼：A文章編號(hào)：1006-0316（2025）06-0001-08

doi：10.3969/j.issn.1006-0316.2025.06.001

Abstract：Traditional image segmentation techniques often rely on deep learning models based on Convolutional Neural Networks and Transformer architectures.Although these models excel at local feature extraction，they have limitations in capturing long-range dependencies.Moreover，such models tend to have a large number of parameters and high computational complexity，which results in significant computational burdens，especially in resource-constrained environments.To address this issue，this paper proposesa lightweight image segmentation method based on Mamba.By introducing Mamba’s eficient architecture combined with the classical U-Net structure， this method aims to tackle the challenges faced by image segmentation models in mobile device scenarios， such as large parameter sizes and inadequate processing speeds for real-time applications.Specificaly，the method incorporates Visual State Space （VSS） blocks，which are used alongside convolutions （CNN） to form hybrid building blocks for capturing extensive contextual information.Additionally，a non-symmetric encoder-decoder structure is designed.Experimentsonthe public dataset ISIC20l7 show that， while maintaining segmentation accuracy，the proposed model reduces the parameter count by 99.94% compared to traditional pure visual Mamba models， by 75.51% compared to the lightest existing visual Mamba U-Net model， and by 99.84% compared to the classic U-Net model. The designed model achieves significant reductions in computational complexity while maintaining excellent segmentation accuracy， thus meeting the demands ofreal-time applications.

Key words ∵ Mamba; image segmentation;lightweight mode;encoder;decoder

隨著深度學(xué)習(xí)在計(jì)算機(jī)視覺(jué)領(lǐng)域的廣泛應(yīng)用，圖像分割技術(shù)在醫(yī)療圖像分析、自動(dòng)駕駛、遙感影像處理等領(lǐng)域得到了越來(lái)越多的關(guān)注和應(yīng)用。（剩余8861字）