We propose Convolutional Block Attention Module (CBAM), a simple yet effective attention module for feed-forward convolutional neural networks. Given an intermediate feature map, our module sequentially infers attention maps along two separate dimensions, channel and spatial, then the attention maps are multiplied to the input feature map for adaptive feature refinement. Because CBAM is a lightweight and general module, it can be integrated into any CNN architectures seamlessly with negligible overheads and is end-to-end trainable along with base CNNs. We validate CBAM through extensive experiments on ImageNet-1K, MS COCO detection, and VOC 2007 detection datasets.

Attention mechanisms have become an integral part of deep neural networks. In the context of CNNs, attention can be broadly categorized as channel attention (what to focus on) and spatial attention (where to focus). Previous works like SENet have explored channel attention alone, demonstrating significant improvements. However, combining both channel and spatial attention in a unified module remains underexplored.

We propose CBAM, which sequentially applies channel and spatial attention modules to any given feature map. The channel attention module exploits the inter-channel relationship of features using both average-pooled and max-pooled features. The spatial attention module utilizes the inter-spatial relationship by applying pooling operations along the channel axis. This sequential arrangement allows the network to first identify informative channels, then locate informative spatial regions within those channels.

SENet (Squeeze-and-Excitation Networks) introduced channel attention via global average pooling followed by fully connected layers. While effective, SENet only considers channel-wise attention and ignores spatial information. STN (Spatial Transformer Networks) introduced spatial attention via learned geometric transformations, but is computationally expensive and difficult to train. Other attention mechanisms in NLP, such as self-attention in Transformers, have demonstrated the power of attention but are not directly applicable to convolutional architectures in their original form.

Given an input feature map F in R^(C x H x W), the Channel Attention Module first generates a channel attention map M_c in R^(C x 1 x 1) by applying both average pooling and max pooling across the spatial dimensions, then passing through a shared multi-layer perceptron (MLP). The outputs are summed and passed through a sigmoid activation: M_c(F) = sigma(MLP(AvgPool(F)) + MLP(MaxPool(F))). Using both pooling operations captures complementary information — average pooling captures general channel statistics while max pooling captures discriminative features.

The Spatial Attention Module takes the channel-refined feature map and generates a spatial attention map M_s in R^(1 x H x W). It first applies average pooling and max pooling along the channel axis, concatenates the results, and then uses a 7x7 convolution layer followed by sigmoid to produce the spatial attention: M_s(F') = sigma(Conv7x7([AvgPool(F'); MaxPool(F')])). The final output is F'' = M_s * (M_c * F), where attention is applied sequentially.

On ImageNet-1K classification, CBAM consistently improves various architectures: ResNet-50 improves from 75.44% to 77.34% top-1 accuracy, MobileNet improves from 68.36% to 70.99%. The overhead is minimal: only 0.06% parameter increase for ResNet-50. Compared to SE module, CBAM achieves better performance with similar computational cost, confirming the benefit of spatial attention.

On MS COCO object detection using Faster R-CNN with ResNet-50 backbone, CBAM improves the mAP from 31.1% to 32.4%. On VOC 2007 detection, the improvement is from 73.2% to 75.1% mAP. Ablation studies confirm that the sequential channel-then-spatial arrangement outperforms parallel or reverse orderings, and that using both average and max pooling is better than using either alone.

We have presented CBAM, a simple and effective attention module that combines channel and spatial attention for convolutional neural networks. Our extensive experiments demonstrate consistent improvements across multiple architectures and tasks with negligible computational overhead. CBAM is a plug-and-play module that enhances CNN feature representations without requiring architectural modifications, making it practical for a wide range of applications.