EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Garment Image Segmentation Method Based on Wavelet Multi-Scale Features and Improved U-Net

Jiangtao Wei, Yanmei Li, Yu Chen

Journal of Design Intelligence · 03.07.2026 · 1 (3) · 25-38

DOI: 10.64509/jdi.13.103 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Design Intelligence, online erschienen am 03.07.2026. Verfasst von Jiangtao Wei und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Against the background of the gradual integration of fashion design and AIGC, obtaining accurate clothing masks through image segmentation and further guiding diffusion models for pattern generation and reconstruction serves as a key technique for intelligent design. Nevertheless, practical scenarios are plagued by issues such as complex background interference and occlusion between human bodies and garments, which easily lead to blurred segmentation boundaries and loss of details, thus impairing the subsequent design generation performance. To address these problems, this paper proposes a clothing image segmentation method based on an improved U-Net architecture. By fusing wavelet sampling and partial convolution in the downsampling stage, the feature learning of the encoder is optimized and the influence of occluded regions is reduced. Furthermore, the Efficient Local Attention-T (ELA-T) mechanism is embedded into the skip connections to enhance the model's sensitivity to subtle boundary features. A novel residual multi-scale depthwise separable convolution module (R-MSDW) is designed to strengthen the model's capability in capturing multi-scale features and hierarchical semantic information. Finally, experiments are conducted on the DeepFashion2 dataset and a self-constructed dataset. The proposed method outperforms mainstream models such as U-Net, SegNet, and U2-Net, achieving an mIoU of 86.1% and an overall accuracy of 95.1% on the DeepFashion2 dataset, which surpasses vanilla U-Net by 8.78% in mIoU and 2.7% in overall accuracy respectively, and an mIoU of 85.4% and an overall accuracy of 94.8% on the self-constructed dataset. The proposed model helps moderately enhance the visual quality of generated fashion design images.

Zitieren

APA

Wei, J., Li, Y., & Chen, Y. (2026). Garment Image Segmentation Method Based on Wavelet Multi-Scale Features and Improved U-Net. Journal of Design Intelligence, 1(3), 25-38. https://doi.org/10.64509/jdi.13.103

Vancouver

Wei J, Li Y, Chen Y. Garment Image Segmentation Method Based on Wavelet Multi-Scale Features and Improved U-Net. Journal of Design Intelligence. 2026;1(3):25-38. doi:10.64509/jdi.13.103

BibTeX
@article{Wei2026,
  author = {Wei, Jiangtao and Li, Yanmei and Chen, Yu},
  title = {Garment Image Segmentation Method Based on Wavelet Multi-Scale Features and Improved U-Net},
  journal = {Journal of Design Intelligence},
  year = {2026},
  volume = {1},
  number = {3},
  pages = {25--38},
  doi = {10.64509/jdi.13.103}
}
Kostenlos starten →