code-MAMNet

MAM: A Multipath Attention Mechanism for Image Recognition

The MXNet implementation of "MAM: A Multipath Attention Mechanism for Image Recognition"

Structure

The overall architecture of MAM. The MAM consists of a channel attention module (CAM) and a spatial attention module (SAM). Channel attention models the correlation between feature channels and improves performance by enhancing essential features. Spatial attention captures the multi-scale information of interest regions in the featuremaps through multi-path.

Requirements

• Python3
• CUDA = 10.1
• MXNet = 1.5.1
• d2lzh
• matplotlib

Result

Classification performance achieved on Downsampled ImageNet. (MAMNet is the CNN network structure embedded with the MAM).

Models	Top-1 acc	Top-5 acc
ResNet50	60.43	82.48
ResNet50-MAM	61.61	84.76

Visualization

Quantitative evaluation results of Grad-CAM. Comparison of visualization results of the MAM integrated network (ResNet50 + MAM) and the baseline (ResNet50). The predicted value represents the softmax score of the network in the ground truth category.

Citation

@article{https://doi.org/10.1049/ipr2.12370,
        author = {Zhang, Hao and Peng, Guoqin and Wu, Zhichao and Gong, Jian and Xu, Dan and Shi, Hongzhen},
        title = {MAM: A multipath attention mechanism for image recognition},
        journal = {IET Image Processing},
        volume = {16},
        number = {3},
        pages = {691-702},
        doi = {https://doi.org/10.1049/ipr2.12370},
        url = {https://ietresearch.onlinelibrary.wiley.com/doi/abs/10.1049/ipr2.12370},
        year = {2022}
}