Journal article 153 views 70 downloads
A2D2C: Adaptive attention-driven dynamic convolution for local feature adaptation
,
Fan Wan,
Xingyu Miao ,
Jingjing Deng ,
Xianghua Xie ,
Yang Long
Pattern Recognition, Start page: 113915
Swansea University Author:
Xianghua Xie
-
PDF | Accepted Manuscript
Author accepted manuscript document released under the terms of a Creative Commons CC-BY licence using the Swansea University Research Publications Policy (rights retention).
Download (10.32MB)
DOI (Published version): 10.1016/j.patcog.2026.113915
Abstract
Dynamic convolution is an advanced deep-learning strategy that enables neural networks to adjust their convolutional kernels dynamically in response to varying input data. This adaptability enhances the network’s efficiency in processing diverse features. However, traditional dynamic convolution tec...
| Published in: | Pattern Recognition |
|---|---|
| ISSN: | 0031-3203 |
| Published: |
Elsevier BV
2026
|
| Online Access: |
Check full text
|
| URI: | https://cronfa.swan.ac.uk/Record/cronfa71874 |
| Abstract: |
Dynamic convolution is an advanced deep-learning strategy that enables neural networks to adjust their convolutional kernels dynamically in response to varying input data. This adaptability enhances the network’s efficiency in processing diverse features. However, traditional dynamic convolution techniques often overlook the critical role of local features in image classification, resulting in suboptimal performance in capturing fine details and textures necessary for accurate image analysis. To address this, our research introduces Adaptive Attention-Driven Dynamic Convolution (A2D2C), an innovative adaptive adjustment mechanism that focuses on local image features, significantly improving the network’s ability to capture fine details and overall performance. Moreover, our paper proposes a novel dynamic convolution that enhances the network’s feature learning ability by combining the input feature map with multiple convolution kernels to generate the attention weights. Additionally, we develop a streamlined version of our model, named A2D2C+, which significantly increases operational efficiency and reduces computational costs. Experimental evaluations on the ImageNet, CIFAR-100 and COCO datasets demonstrate substantial performance enhancements, underscoring the efficacy and applicability of our approach. |
|---|---|
| Keywords: |
Attention; Dynamic convolution; Local features |
| College: |
Faculty of Science and Engineering |
| Start Page: |
113915 |