.png){kind=icon}
.png){kind=icon}
Return
Multi-source Feature Map Distillation for enhanced low-resolution object recognition
.png){kind=icon}
PRE
AI
DOI：10.1016/j.compeleceng.2025.110710.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Knowledge distillation is an effective method for addressing the problem of low-resolution object recognition. However, due to resolution differences, the feature map sizes become inconsistent, making it difficult for the student model to fully learn the rich privileged information contained in the teacher model. Our previous work addressed this issue through a feature decoder, achieving cross-resolution feature map distillation. However, it fails to fully leverage both high-resolution samples and their feature maps to extract more privileged information at the distillation point. To this end, this paper proposes a Multi-source Feature Map Distillation (MsFMD) method to further improve the performance of low-resolution object recognition in practical applications such as intelligent video surveillance. We design a feature decoder with a channel attention mechanism to better leverage the privileged information from the teacher model and employ multi-level decoder modules to process deep features, achieving multi-level feature map distillation. Additionally, this paper introduces a joint data augmentation method, effectively enhancing the student model’s adaptability and robustness across samples with varying resolutions. The overall performance of MsFMD is verified in multiple recognition tasks by comparing it with state-of-the-art knowledge distillation methods on low-resolution and noisy objects.
Journal
C
IF:
4.9
Papers:
6.7K
Citations:
1.3W