.png){kind=icon}
.png){kind=icon}
Return
AdaptiveMamba: Comprehensive visual representation learning with adaptive semantic perception
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2025.115031.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
State Space Models (SSMs) have proven effective in visual tasks, which capture features through sequential scanning with near-linear complexity. However, their scanning-based learning limits the ability to model relationships between distant patches and distinguish key semantic features within an image, thereby hindering global information capture and image understanding. To address these challenges, we propose AdaptiveMamba, which enhances the capturing of global information and the attention to the key semantic features. AdaptiveMamba incorporates the Self-guided Semantic Perception (SSP) block, designed to adaptively highlight salient semantic features and enhance the extraction of global information. Specifically, in the SSP block, we adopt the semantic guidance module to dynamically score the semantic importance of each patch and use the comprehensive perception module to extract both salient and global semantic features effectively. Extensive experiments and analyses demonstrate that AdaptiveMamba consistently achieves superior performance across various visual tasks, attaining 84.7% top-1 accuracy on ImageNet, 49.5 % APb on COCO, and 51.1 % mIoU on ADE20K, while maintaining a lower computational cost. Furthermore, the integration of the SSP block into diverse architectures showcases its strong generality, achieving a 3.8% accuracy improvement on ResNet-101 while also producing varying degrees of enhancement across other models, thereby underscoring the critical role of adaptive semantic perception in visual representation learning. Our code is available at: https://github.com/GZU-SAMLab/AdaptiveMamba.
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W