R-CBAM: Residual Convolutional Attention Module

• R-CBAM is a design family that embeds sequential channel and spatial attention within residual blocks to adaptively refine CNN feature maps.
• It integrates standard CBAM into two 3x3 convolutional operations with residual connections, enhancing feature extraction in tasks like deraining and classification.
• Empirical evidence shows that R-CBAM improves metrics such as PSNR and accuracy, underlining its practical impact in modern deep convolutional networks.

Searching arXiv for papers on “Residual Convolutional Block Attention Module” and CBAM in residual networks. First search: exact phrase and acronym. Residual Convolutional Block Attention Module (R-CBAM) denotes a class of CNN components that combine residual learning with sequential channel and spatial attention. The term is not historically uniform. The original CBAM work established the standard attention mechanism and integrated it into ResNet blocks by applying CBAM to “the convolution outputs in each block,” but it did not use the name R-CBAM (Woo et al., 2018). Subsequent literature used closely related formulations, including a “residual CBAM refinement network” for light-field view synthesis (Gul et al., 2020) and a CBAM-augmented ResNet50 for ship classification that is R-CBAM-like in interpretation but not formally specified as a new block (Kwon et al., 2024). A formally named “Residual Convolutional Block Attention Module (R-CBAM) Block” appears in a U-Net-based deraining network in which residual convolution, channel attention, spatial attention, and skip addition are defined as a single encoder–decoder building block (Si et al., 31 Jul 2025). Taken together, this suggests that R-CBAM is better treated as a residual-attention design family than as a single universally fixed operator.

1. Terminological status and historical emergence

The conceptual origin of R-CBAM lies in CBAM itself. CBAM introduced a lightweight attention module that sequentially infers channel and spatial attention maps and multiplies them with an intermediate feature map for adaptive refinement; in residual networks, it is inserted on the convolution outputs inside each block (Woo et al., 2018). This established the canonical residual-attention pattern from which later R-CBAM-style designs were derived.

Later papers adopted this residualized use of CBAM with different levels of formalization. In light-field view synthesis, the refinement stage is explicitly described as using a “residual convolutional block attention module (CBAM),” but the attention equations remain standard CBAM and the novelty lies in the residual refinement architecture rather than in a new attention law (Gul et al., 2020). In optical satellite ship classification, CBAM is integrated into a ResNet50 backbone and then combined with multi-scale feature integration, depthwise separable convolutions, and dilated convolutions; however, the paper does not define a standalone residual-attention operator or specify the exact insertion granularity, so it is best treated as a CBAM-augmented residual network rather than a formally defined R-CBAM block (Kwon et al., 2024).

A more explicit block-level use appears in the deraining literature, where R-CBAM is named and described as a module composed of two $3\times3$ convolutions, SiLU activations, channel attention, spatial attention, and a residual/skip connection, and is inserted into both encoder and decoder stages of a U-Net-like network (Si et al., 31 Jul 2025). By contrast, some later residual U-Net papers use CBAM on skip features rather than within each residual unit, which is relevant to R-CBAM only in a looser architectural sense (Yuan et al., 23 May 2025).

Paper	Residual-CBAM usage	Formal R-CBAM status
(Woo et al., 2018)	CBAM on convolution outputs in each ResBlock	No
(Gul et al., 2020)	Residual CBAM refinement network	No new attention formula
(Kwon et al., 2024)	CBAM-augmented ResNet50	R-CBAM-like, not formal
(Si et al., 31 Jul 2025)	Encoder/decoder R-CBAM Block	Yes
(Yuan et al., 23 May 2025)	CBAM on Res-UNet skip connections	Not a reusable residual block

2. Canonical block anatomy

In its most explicit formulation, R-CBAM is a residual convolutional block whose internal transform is refined by sequential channel and spatial attention before skip addition. The deraining formulation defines the block around an input feature map $X$, two consecutive $3\times3$ convolutions, channel attention built from global average pooling and global max pooling passed through a shared MLP, spatial attention built from channel-wise average and max projections followed by a $7\times7$ convolution, and a residual output that adds either $X$ directly or a $1\times1$ projection of $X$ when channel dimensions differ (Si et al., 31 Jul 2025).

This distinguishes R-CBAM from both of its immediate antecedents. Relative to a plain residual block, R-CBAM inserts data-dependent channel-wise and spatial-wise gating between the convolutional transform and the residual addition. Relative to a standard CBAM, it is not merely an attention plug-in attached somewhere in a host network; rather, CBAM is embedded inside a two-convolution residual unit and therefore participates directly in the block’s internal residual mapping (Si et al., 31 Jul 2025).

The light-field formulation adopts the same high-level organization with slightly different local details. Each CBAM block in that refinement network has two $3\times3$ convolutions with ReLU activation in between, followed by channel attention, spatial attention, and residual addition; the block is then repeated inside a larger residual refinement stage (Gul et al., 2020). The original CBAM paper supplies the standard attention internals—channel attention first, then spatial attention—and shows how this sequence can be inserted into residual architectures without redesigning the identity path (Woo et al., 2018).

One implementation ambiguity is worth noting because it recurs in later reproductions. In the 2025 deraining paper, the text says that the block consists of two consecutive $3\times3$ convolutional layers followed by the SiLU activation function, but the displayed equation places SiLU only after the first convolution. The safest reading is therefore $$\mathrm{Conv}_{3\times3}\rightarrow \mathrm{SiLU}\rightarrow \mathrm{Conv}_{3\times3}$$, with no explicitly specified second SiLU before attention (Si et al., 31 Jul 2025).

3. Mathematical formulation

The standard CBAM core is the sequential refinement

$X$0

where $X$1 is channel attention, $X$2 is spatial attention, and $X$3 denotes element-wise multiplication (Woo et al., 2018). In CBAM, channel attention is computed as

$X$4

and spatial attention is computed as

$X$5

This channel-first ordering outperformed spatial-first and parallel variants in the original ablations on ResNet-50, and thus became the default CBAM sequence (Woo et al., 2018).

The formally named R-CBAM block in deraining specializes this pattern by inserting it inside a residual two-convolution block: $X$6

$X$7

$X$8

$X$9

$3\times3$0

$3\times3$1

Within this formulation, $3\times3$2, $3\times3$3, and the block preserves spatial resolution; if channel counts differ, only the skip path is projected (Si et al., 31 Jul 2025).

For residual backbones that use standard CBAM but do not formalize R-CBAM as a separate block, the most defensible reconstruction is

$3\times3$4

because CBAM is applied to the convolution outputs in each block rather than to the identity path or explicitly to the post-addition tensor (Woo et al., 2018). This interpretation is strongly supported by the residual-block integration description, but it remains an interpretation rather than an independently named operator.

4. Placement patterns across architectures

R-CBAM-style integration is highly architecture-dependent. In the original CBAM study, the module is inserted into ResNet residual blocks by operating on the convolution outputs before shortcut aggregation, thereby refining the residual branch while leaving the identity path unchanged (Woo et al., 2018). This is the cleanest residual-block placement and remains the reference design for residualized CBAM.

In light-field view synthesis, the residual CBAM component appears only in the third stage of a three-stage pipeline: stereo feature extraction, disparity estimation, and final image refinement. The refinement network has a head $3\times3$5 convolution, five residual groups, three CBAMs per group, a tail $3\times3$6 convolution, and a final residual correction added to the average of the warped views. Here “residual” applies both inside each CBAM block and at the output of the whole refinement stage (Gul et al., 2020).

In the deraining network, R-CBAM is used more symmetrically. All four encoder blocks are composed of an R-CBAM block, and decoder stages also apply R-CBAM after upsampling and skip concatenation so that merged low-level and high-level features are further refined before reconstruction (Si et al., 31 Jul 2025). This encoder–decoder deployment makes the block simultaneously responsible for feature abstraction and feature recovery.

Not every residual architecture that uses CBAM constitutes an R-CBAM in this strict sense. The ship-classification model integrates CBAM into ResNet50 and later augments the architecture with multi-scale feature integration, depthwise separable convolutions, and dilated convolutions, but the paper does not specify whether CBAM is inserted in every bottleneck, after selected blocks, after each stage, or near the classifier head (Kwon et al., 2024). Likewise, the neonatal brain-ultrasound segmentation network is a Residual U-Net in which CBAM is applied to encoder features before skip transfer to the decoder and then combined with a Sparse Attention Layer; this is best described as CBAM-enhanced skip connections within a residual U-Net rather than as a canonical residual-attention block (Yuan et al., 23 May 2025).

5. Empirical evidence across domains

The empirical case for residualized CBAM begins with the original CBAM benchmarks. On ImageNet-1K, ResNet50 improved from Top-1/Top-5 error $3\times3$7 to $3\times3$8 when CBAM was inserted into residual blocks, and ResNet101 improved from $3\times3$9 to $7\times7$0. On MS COCO detection with Faster R-CNN, CBAM raised mAP@[.5,.95] from $7\times7$1 to $7\times7$2 for a ResNet-50 backbone and from $7\times7$3 to $7\times7$4 for ResNet-101 (Woo et al., 2018). These results established that residual backbones benefit measurably from channel-then-spatial refinement.

Later domain-specific studies show that the same design principle transfers beyond classification. In light-field view synthesis, the attention-enabled residual refinement network improved over the identical refinement stage without attention: on the Diverse dataset, PSNR/MS-SSIM increased from $7\times7$5 to $7\times7$6, and on Flowers from $7\times7$7 to $7\times7$8. The paper’s qualitative analysis attributes the gain primarily to better handling of occlusions and difficult local reconstructions (Gul et al., 2020).

In optical satellite ship classification, a baseline ResNet50 achieved accuracy $7\times7$9, ResNet50 with standard CBAM achieved $X$0, and an enhanced ResNet model with improved CBAM together with multi-scale feature integration, depthwise separable convolutions, and dilated convolutions reached $X$1. Class-wise metrics also improved substantially, with oil tanker precision/recall/F1 rising to $X$2, and attention heatmaps showing more concentrated activation over ship structures under varying background complexity (Kwon et al., 2024).

The formally named R-CBAM block in deraining is associated with strong restoration performance, although its contribution is not isolated by architecture-only ablation. The full model reports $X$3 PSNR and $X$4 SSIM on Rain100L and $X$5 PSNR and $X$6 SSIM on Rain100H, and the discussion ties the gains to dynamic emphasis on rain-affected regions, boundaries, and textures through channel and spatial attention (Si et al., 31 Jul 2025).

Residual U-Net variants that place CBAM outside the residual block proper still provide useful comparative evidence. In brain ultrasound segmentation, a Res-UNet reached Dice/IoU $X$7, adding CBAM to the skip connections raised performance to $X$8, and combining CBAM with Sparse Attention Layer yielded $X$9. These numbers indicate that CBAM-style refinement remains effective even when used as skip-feature attention rather than as a self-contained residual block (Yuan et al., 23 May 2025).

6. Conceptual limits, common misconceptions, and successor directions

The main misconception surrounding R-CBAM is that the term denotes a universally standardized module. The literature does not support that reading. Some papers provide a genuine residual-attention block with explicit convolution, attention, and skip-add sequencing; others merely use CBAM in a residual backbone; still others place CBAM on skip connections or late refinement stages without defining a reusable residual-attention primitive. The ship-classification paper is explicit that it is useful as a reference for CBAM used in a residual network but not as a precise source for a newly defined “Residual Convolutional Block Attention Module” with exact block wiring (Kwon et al., 2024).

Even papers that formally name R-CBAM may leave important reproducibility gaps. The deraining paper does not provide a direct ablation comparing U-Net without R-CBAM, U-Net with ordinary residual blocks, U-Net with plain CBAM, and U-Net with R-CBAM; its ablation isolates only the loss terms. It also omits some R-CBAM-specific hyperparameters, including channel widths at each U-Net stage, the channel-attention reduction ratio, padding and stride settings, and normalization choices (Si et al., 31 Jul 2025). The light-field paper likewise uses standard CBAM mechanics inside a residual refinement architecture rather than introducing a new attention formula (Gul et al., 2020).

This lack of standardization has encouraged successor modules that preserve residual compatibility while modifying CBAM’s internals. NAM retains CBAM-like block insertion but derives attention from normalization coefficients, reporting on ResNet50/CIFAR-100 a Top-1 error of $1\times1$0 for channel-only NAM versus $1\times1$1 for CBAM, with substantially lower channel-attention parameter overhead (Liu et al., 2021). GAM keeps the channel-then-spatial sequence but replaces CBAM’s pooled descriptors with a 3D-permutation MLP channel branch and a two-convolution spatial branch; on ImageNet-1K with ResNet50 it reports Top-1 error $1\times1$2 versus $1\times1$3 for CBAM, albeit with much larger parameter and FLOP cost (Liu et al., 2021). RFAConv goes further by arguing that ordinary CBAM-style spatial attention acts on the wrong object—feature-map positions rather than receptive-field features—and proposes receptive-field-aware variants such as RFCBAMConv, which on ResNet18 reaches Top-1 $1\times1$4 versus $1\times1$5 for CBAMConv (Zhang et al., 2023).

Accordingly, R-CBAM occupies a specific place in the evolution of CNN attention: it is the residualized form of CBAM’s channel–spatial refinement, but it is also a moving target whose exact realization depends on where attention is inserted, whether the residual connection wraps the attention-refined transform or merely accompanies it elsewhere in the architecture, and whether later work preserves or replaces the standard CBAM channel and spatial submodules.