Home / Curated / Layers & Modules / Balanced Attention Mechanism

Balanced Attention Mechanism

Layers & Modules AIMech:BalancedAttentionMechanism DRAFT

Aliases: BAM (balanced attention)

A convolutional feature-gating component for single-image super-resolution that computes average-pooled channel attention and maximum-pooled spatial attention in parallel, multiplies their sigmoid gates by broadcasting, and scales the input feature tensor.

Identity Notes & Concept Scope Compared all 442 pinned AIO and 595 native IDs, names, aliases and scoped synonyms; no normalized identity match. Ignored- and hidden-inclusive fixed-string searches of the recovery, older discovery and original Documents checkouts returned 166, 106, 105 lines. All distinct match contents were inspected. Semantic matching covered 66 AIO and 27 native definitions, with every matched definition read. 7 full annotated AIO node/edge/generated contexts and 3 full native records were inspected. The selected identity is Wang and colleagues’ 2021 parallel average-channel/max-spatial product gate, not Park and colleagues’ 2018 Bottleneck Attention Module. BAM is qualified in the alias. The ACAM/MSAM branches are described within this composite layer, not asserted to be equivalent to existing pooling or activation classes. CBAM is sequential and SE has only channel recalibration. The inspected generic AIO Layer is broader; no exact counterpart was found.

Claims & Grounded Evidence

6 Claims

Mechanism #average-channel-branch

For input F with batch, channel, height and width axes, the ACAM branch averages over spatial positions separately for each channel. It applies a learned 1-by-1 channel reduction, PReLU, a 1-by-1 expansion to the original channel count, and sigmoid. Its channel weights are shared over spatial positions. The text uses reduction ratio 16 depending on the channel count; this is a configuration choice, not a universal identity requirement.

supports Source: BAM: A Balanced Attention Mechanism for Single Image Super Resolution · Locator: Proposed method: Avgpool Channel Attention Module, PDF pp. 2–3; Equations 1 and 3 and Figure 2, p. 3

Mechanism #maximum-spatial-branch

The MSAM branch starts from the same unmodified F, takes the maximum across channels at every spatial position, and applies a learned 7-by-7 convolution followed by sigmoid to produce one spatial gate shared across channels. This maximum is over the channel axis, whereas ACAM averages over spatial axes. Neither branch consumes the output of the other in the paper’s parallel construction.

supports Source: BAM: A Balanced Attention Mechanism for Single Image Super Resolution · Locator: Proposed method: Maxpool Spatial Attention Module, PDF pp. 3–4; Equation 2 and Figure 2, p. 3; Equation 4, p. 4

Scope #product-gate-boundary

Broadcast the channel and spatial sigmoid outputs, multiply them together, and multiply the resulting gate into F. The core equation contains no residual addition or softmax normalization. For finite real logits the factor lies strictly between zero and one, so it attenuates immediate nonzero feature magnitudes; numerical sigmoid saturation can reach endpoints. For each batch item, the channel-by-flattened-spatial gate is an outer product of rank at most one, not an arbitrary gate at every channel-position pair. These are direct consequences of the formula, not guarantees of image denoising.

supports Source: BAM: A Balanced Attention Mechanism for Single Image Super Resolution · Locator: Figure 2, p. 3; Balanced Attention Mechanism, p. 4, Equation 5; direct algebra and constructed rational gate examples

Limitation #gradient-and-ablation-limits

The paper motivates its parallel branches as balancing noise suppression and texture preservation and writes a product-rule gradient expansion. Shared gradient paths alone do not prove convergence, optimal balance or separate causal contributions of each branch. Table 2 compares complete attention modules inside IMDN; it does not provide an isolated ACAM-versus-MSAM ablation that establishes those individual effects.

supports Source: BAM: A Balanced Attention Mechanism for Single Image Super Resolution · Locator: Balanced Attention Mechanism, p. 4, Equation 6; Ablation Study and Table 2, p. 7

Scope #evaluation-protocol

The reported training uses DIV2K training images with augmentation, 64-by-64 low-resolution patches, batch size 16, L1 loss and Adam for 1000 epochs. Evaluation uses PSNR and SSIM on the Y channel with bicubic downsampling at factors 2, 3 and 4. The paper compares several surrounding super-resolution architectures and notes that some original baselines use extra training data, so the table is not evidence of uniformly matched retraining or architecture-independent gains.

supports Source: BAM: A Balanced Attention Mechanism for Single Image Super Resolution · Locator: Experiments: Datasets and metrics, Implementation details and Comparison experiments, PDF p. 4; Table 1, p. 5

Limitation #mixed-results-and-runtime

The reported results are metric-, scale-, dataset- and architecture-specific. Table 1 shows PAN on Manga109 at factor 3 with PSNR decreasing from 33.61 to 33.19 while SSIM increases from 0.9448 to 0.9453; RCAN at factor 4 on Manga109 also has lower PSNR after replacement. The speed comparison in Figure 6 uses 700 images per input scale on an RTX 2080 Ti for IMDN and DRLN variants. It does not isolate parallelism as the sole cause or establish universal accuracy or runtime improvements.

supports Source: BAM: A Balanced Attention Mechanism for Single Image Super Resolution · Locator: Table 1, PDF p. 5; Speed comparison and Figure 6, p. 7

⚠️ Curated Limitations & Negative Results

The canonical identity is the 2021 balanced attention component, not Park and colleagues’ 2018 Bottleneck Attention Module or a complete super-resolution network. The shared acronym BAM is qualified in the alias.
This record follows the eight-page arXiv v3 paper. The author code was not inspected or executed, so padding, bias, small-channel rounding, trained checkpoints and implementation fidelity are not certified.
The paper PDF retains template author and affiliation placeholders and a 2022 copyright footer. Authors and the 2021 version date come from the primary arXiv metadata; no conference acceptance is inferred.
Chosen exact gates illustrate algebraic limits only. No model was trained, no image or hardware benchmark was replicated, and the record remains DRAFT.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:Layer` Layer	broad_match	The pinned Layer class is broader than this learned input-to-output feature-map component. The pinned AttentionLayer denotes dot-product attention, which is not the channel/spatial gating procedure here; pooling, convolution and activation targets describe constituent operations rather than equivalent whole-module identities.

Evidence Sources & Primary Literature

1 Source

Source ID	Publication / Document	Type / Access	Reference & Checksum
`wang2021`	BAM: A Balanced Attention Mechanism for Single Image Super Resolution Fanyi Wang, Haotian Hu, Cheng Shen	primary paper full_text	ID: https://arxiv.org/abs/2104.07566v3 sha256:dcd27007bcb0…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-13T09:02:46Z

Updated: 2026-10-13T09:11:02Z

Source File: data/curated/BalancedAttentionMechanism.yaml