Collaborative mining enables efficient multi-scale feature compression for machine vision
Machine-oriented multi-scale feature compression aims to remove task-insensitive redundancy while preserving task-critical semantic and structural information. Existing methods primarily focus on reducing cross-scale redundancy, while intra-scale redundancy within individual feature maps remains insufficiently explored. To address this limitation, we propose a Collaborative Mining Network (CMNet), a unified redundancy-aware framework that progressively suppresses both intra-scale and cross-scale redundancies. CMNet first employs an intra-scale redundancy-aware module to suppress redundant responses within each feature map while preserving task-relevant information. It then uses a context-guided adaptive fusion module to dynamically model spatial and channel correlations across scales, thereby reducing cross-scale redundancy. Experiments on multiple downstream machine vision tasks demonstrate that CMNet achieves superior rate-accuracy performance over existing multi-scale feature compression methods, enabling more efficient transmission of task-relevant multi-scale features.
Authors
- Zhaoqing Pan (ORCID: https://orcid.org/0000-0003-1390-399X)
- Yun Zhang
- Renyi Chen (ORCID: https://orcid.org/0000-0002-6564-2376)
- Fu Lee Wang
- Haoran Xie
- Qiang Liu
Institutions
- Sun Yat-sen University (CN)
- Tianjin University (CN)
- Lingnan University (HK)
- Hong Kong Metropolitan University (HK)
Publication Details
- Journal
- Communications Engineering
- Published
- 2026-11-20
- DOI
- https://doi.org/10.1038/s44172-026-00822-1
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- National Natural Science Foundation of China