EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC 4.0

Multi-scale Feature Extraction and Dynamic Aggregation for Referring Image Segmentation

Xinzhuo Gao, Quange Tan, Siyu Meng, Mingjie Wang, Rong Wang

Frontiers in Computing and Intelligent Systems · 04.03.2026 · 15 (2) · 46-56

DOI: 10.54097/emt8dy47 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Frontiers in Computing and Intelligent Systems, online erschienen am 04.03.2026. Verfasst von Xinzhuo Gao und 4 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Image referring segmentation aims to accurately locate and segment the corresponding target regions in an image based on textual descriptions. Existing methods mainly face three major challenges: Firstly, there is insufficient information interaction in the extraction process of multi-scale visual features, making it dif-ficult to balance local details and global semantics. Secondly, cross-modal semantic alignment lacks a bidirectional guidance mechanism, resulting in weak semantic consistency between image and text features. Thirdly, the feature aggregation strategy lacks dynamic adaptability and cannot flexibly adjust feature weights according to scene complexity, leading to insufficient feature discriminability. To address the above issues, a referring image segmentation method based on a chained collaborative mechanism for multi-scale feature extraction and dynamic aggregation is proposed. Firstly, a multi-scale adaptive feature fusion module is constructed. The image features after patch embedding are split into four channel sub-spaces through group convolution. After extracting multi-scale features via adaptive max-pooling, the gating network dynamically adjusts the fusion weights to realize the adaptive interaction between local details and global semantics, thus improving the diversity of feature expression. Secondly, a dual-attention modal alignment module is designed. Channel-first convolutional attention is introduced into image fea-tures, which adjusts the weights of different channels and spatial positions in the feature map under the guidance of text features to strengthen the expression of key regions. Efficient channel attention is in-troduced into text features, which takes image features as prior and enhances the semantic relevance of text through channel interaction, improving the accuracy and efficiency of feature alignment. Finally, a dy-namic feature aggregation module is proposed to explore the non-linear dependencies between channels, generate fine-grained feature weights, and dynamically adjust features of each scale through competitive weight allocation, so as to realize the complementary aggregation of global and local features and improve the discriminability of features. Experimental results demonstrate that the proposed method achieves In-tersection over Union (IoU) scores of 74.34% and 66.59% on testA and testB of the RefCOCO dataset, 68.99% and 53.87% on testA and testB of the RefCOCO+ dataset, and 61.65% on the test set of the G-Ref dataset, respectively, verifying the effectiveness of the proposed method.

Zitieren

APA

Gao, X., Tan, Q., Meng, S., Wang, M., & Wang, R. (2026). Multi-scale Feature Extraction and Dynamic Aggregation for Referring Image Segmentation. Frontiers in Computing and Intelligent Systems, 15(2), 46-56. https://doi.org/10.54097/emt8dy47

Vancouver

Gao X, Tan Q, Meng S, Wang M, Wang R. Multi-scale Feature Extraction and Dynamic Aggregation for Referring Image Segmentation. Frontiers in Computing and Intelligent Systems. 2026;15(2):46-56. doi:10.54097/emt8dy47

BibTeX
@article{Gao2026,
  author = {Gao, Xinzhuo and Tan, Quange and Meng, Siyu and Wang, Mingjie and Wang, Rong},
  title = {Multi-scale Feature Extraction and Dynamic Aggregation for Referring Image Segmentation},
  journal = {Frontiers in Computing and Intelligent Systems},
  year = {2026},
  volume = {15},
  number = {2},
  pages = {46--56},
  doi = {10.54097/emt8dy47}
}
Kostenlos starten →