- ResNet-18 accuracy
- 94.68%
- MobileNetV2 accuracy
- 93.10%
Research question
Channel attention gives CNNs a meaningful accuracy boost for very little compute. But how do different attention designs actually trade accuracy against efficiency, and can a lighter module keep the gains?
In plain language
Attention modules help a network decide which feature channels matter. Squeeze-and-Excitation (SE) and Efficient Channel Attention (ECA) are two popular designs. This study compares them with a new, lighter module, Lite Channel Attention (LCA), on two common architectures, measuring not just accuracy but parameters, FLOPs and real GPU latency.
Method
LCA uses adaptive one-dimensional convolutions with grouped operations to reduce parameter usage while preserving effective attention behavior. All three modules were added to ResNet-18 and MobileNetV2 and trained on CIFAR-10 under the same setup.
Results
LCA reaches competitive accuracy, 94.68% on ResNet-18 and 93.10% on MobileNetV2, while matching ECA in parameter efficiency and keeping favorable inference latency. The paper includes full benchmarks of FLOPs, parameter counts and GPU latency, with practical guidance for deploying attention-enhanced CNNs in resource-constrained environments.
Cite
@article{kanaparthi2026lightweight,
title = {Lightweight Channel Attention for Efficient CNNs},
author = {Kanaparthi, Prem Babu and Padamata, Tulasi Venkata Sri Varshini},
journal = {arXiv preprint arXiv:2601.01002},
year = {2026}
}