PAAM (Parameter-free Attentional Aggregation Model)

The channel attention mechanism and spatial attention mechanism are crucial in enhancing the performance of convolutional neural networks. However, most existing methods focus on developing more intricate attention modules to improve performance, which inevitably increases the number of model parameters. To address the trade-off between performance and parameter count, this paper introduces an efficient Parameter-free Attention Aggregation Model (PAAM) plug-and-play module. The module first creates a Local Feature Enhancement Module (LFEM) using adaptive pooling. Firstly, the local feature enhancement module (LFEM) is constructed through adaptive pooling to enhance the expression of local features; secondly, the local-global feature interaction module (L-GFIM) is used to realize the mutual compensation between local and global features, which effectively extends the coverage of local-global interaction. The experimental results indicate that PAAM outperforms the SOTA model in ImageNet-1K, Cifar-10, and Cifar-100 image classification datasets.