ECA-DenseNet: an image classification model based on efficient channel attention

Abstract

With the rapid development of deep learning, Convolutional Neural Networks (CNNs) have achieved remarkable performance in image classification. However, as network depth increases, issues of high computational complexity and memory consumption become prominent, especially on resource‑limited devices. As a novel CNN architecture, DenseNet enhances feature propagation and reuse via dense connections, yet its excessive computational and memory cost restricts deployment on edge devices. Accordingly, optimizing DenseNet for better efficiency and performance has become an important research direction. This paper proposes an improved DenseNet model integrated with the Efficient Channel Attention (ECA) mechanism, termed ECA‑DenseNet. By embedding the lightweight ECA module, the network adaptively emphasizes informative features and further boosts classification performance. Experiments are conducted on the CIFAR‑10 dataset using standard training and testing pipelines, with the SGD optimizer and cross‑entropy loss. Results demonstrate that ECA‑DenseNet maintains high accuracy while introducing negligible extra overhead, thus achieving better efficiency and lower memory usage, showing clear advantages over vanilla DenseNet. This confirms that the lightweight channel attention mechanism is well suited for optimizing DenseNet, particularly in computation‑constrained scenarios.