Efficient Pyramid Channel Attention (EPCA)

• The paper introduces efficient multi-scale channel attention using pyramid strategies to recalibrate feature maps with minimal parameter overhead.
• It employs grouped convolutions with varying kernel sizes and pooling-fusion techniques to adaptively capture global and local channel descriptors.
• EPCA modules enable plug-and-play integration for both general vision and medical imaging, facilitating parameter-efficient transfer learning.

Efficient Pyramid Channel Attention (EPCA) refers to a class of lightweight, multi-scale attention modules designed to enhance channel context modeling within convolutional neural networks (CNNs). EPCA mechanisms extract and fuse global and local channel-wise descriptors across multiple contexts using pyramid strategies, enabling effective recalibration of feature maps with minimal increases in parameters and computation. These modules have been instantiated and evaluated in both general-purpose computer vision (notably as the core of EPSANet (Zhang et al., 2021)) and specialized medical imaging tasks (such as pathological myopia recognition (Zhang et al., 2023)), demonstrating strong performance and parameter-efficiency compared to established channel attention architectures.

1. Architectural Designs of EPCA

Two primary forms of EPCA have been proposed: (1) Pyramid Squeeze Attention (PSA), forming the core of the EPSA block used in EPSANet, and (2) a simplified multi-scale pooling and channel fusion mechanism, as in the EPCA module for pathological myopia recognition.

EPSA/PSA-based EPCA (Zhang et al., 2021):

• Input tensor $X\in\mathbb{R}^{C\times H\times W}$ is split into $S$ groups along the channel axis (with $C'=C/S$).
• Each branch applies a $k_i \times k_i$ grouped convolution ($k_i=2(i+1)+1$, $G_i=2^{(k_i-1)/2}$), producing $F_i\in\mathbb{R}^{C'\times H\times W}$.
• Each $F_i$ is squeezed via global average pooling to $g_{i,c}$, passed through a two-layer excitation (SEWeight): $Z_i=\sigma(W_1\cdot \mathrm{ReLU}(W_0g_i))$.
• Scale-wise channel statistics $S$0 are normalized across scales via a softmax for each channel: $S$1.
• Outputs are reweighted and concatenated: $S$2, where $S$3.

Pooling-Fusion EPCA (Zhang et al., 2023):

• For input $S$4, multiple average-pooling descriptors $S$5 are computed over different bin sizes (e.g., global $S$6 and local $S$7 grid), where $S$8.
• All $S$9 descriptors concatenated to $C'=C/S$0.
• For each channel, a linear scale-context feature mapping (SCFM) fuses the $C'=C/S$1 descriptors with learnable weights $C'=C/S$2: $C'=C/S$3.
• The channel attention weights are $C'=C/S$4 and applied to scale the input feature map $C'=C/S$5.

2. Integration into Backbone Networks

EPSANet (General Vision) (Zhang et al., 2021):

• The EPSA block replaces the standard $C'=C/S$6 convolution in a ResNet bottleneck (sequence: $C'=C/S$7).
• The residual connection is preserved; if dimension change is required, a $C'=C/S$8 convolution is applied to the skip path.
• Multiple EPSA blocks are stacked to construct EPSANet architectures (e.g., EPSANet-50, EPSANet-101).
• Two main variants are constructed: "Small" (mid-channels $C'=C/S$9, group schedule $k_i \times k_i$0) and "Large" (mid-channels $k_i \times k_i$1, group size $k_i \times k_i$2).

EPCA-Net (Medical Imaging) (Zhang et al., 2023):

• EPCA modules are appended after the last convolutional layer of every residual block (basic or bottleneck) in ResNet architectures.
• For ResNet-18, EPCA is attached after each pair of $k_i \times k_i$3 convolutions; for ResNet-50, after the $k_i \times k_i$4 bottleneck.
• The resulting network, EPCA-Net, retains all residual connections and uses EPCA as a lightweight recalibration plug-in.

3. Multi-Scale Contextual Modeling

EPCA modules incorporate multi-scale context via the following strategies:

• Grouped convolutions with multiple kernel sizes: EPSA/PSA branches apply grouped convolutions of increasing kernel size ($k_i \times k_i$5) to leverage receptive fields of varying spatial extent (Zhang et al., 2021).
• Pyramid pooling: The pooling-fusion EPCA uses average pooling over both global ($k_i \times k_i$6) and local ($k_i \times k_i$7 grid) windows, yielding fine-grained spatial descriptors tuned to clinical priors in pathological myopia recognition (Zhang et al., 2023).

The combination of descriptors is performed in a channel-dependent but scale-fused manner, using a softmax (EPSA) or learned linear weighting (EPCA).

4. Computational Efficiency and Parameter Analysis

EPCA blocks exhibit minimal growth in parameters and FLOPs when integrated into deep networks.

• Parameter and FLOP impact in ImageNet-scale networks (Zhang et al., 2021):

Backbone	Params (M)	FLOPs (G)	Top-1 Acc (%)
ResNet-50	25.56	4.12	75.20
SENet-50	28.07	4.13	76.71
EPSANet-50 (Small)	22.56	3.62	77.49
EPSANet-50 (Large)	27.90	4.72	78.64

• Parameter efficiency in ResNet-18/50 for EPCA (Zhang et al., 2023):
  • ResNet-18: 11.178M → 11.178M (+0)
  • ResNet-50: 23.512M → 23.514M (+0.002M)
  • GFLOPs increase negligibly ($k_i \times k_i$80.001)
  • SE, SPA, and CBAM adapters increase parameter counts by 0.09M–2.5M per network.

EPCA’s architectural design, particularly the use of channel grouping and context pooling, constrains parameter increases to $k_i \times k_i$9 per block.

5. Empirical Performance and Representative Results

ImageNet and COCO (EPSANet, (Zhang et al., 2021)):

• On ImageNet, EPSANet-50 attains 1.93% higher Top-1 accuracy over SENet-50.
• Object detection (Faster-RCNN, Mask-RCNN, RetinaNet) on COCO shows AP gains of up to +2.7 over SENet-50; mask AP improvement of +1.7.
• Ablation reveals that group size scheduling ($k_i=2(i+1)+1$0) and multi-scale kernel design yield best trade-offs; concatenation across branches maintains feature diversity.

Pathological Myopia Recognition (EPCA-Net, (Zhang et al., 2023)):

• EPCA-Net50 achieves 97.56% accuracy on PM-Fundus, surpassing ViT (94.71%), SE (96.61%), and CBAM (96.34%).
• As an adapter (with ResNet frozen), EPCA yields 96.34% accuracy with only 0.002M tunable parameters, outperforming SE- and CBAM-adapters (~2.5M parameters).
• Five-fold cross-validation demonstrates robustness (90.42% ± 4.62% accuracy).
• Receptive field ablations confirm that combining global and $k_i=2(i+1)+1$1 local bins ({1,3}) is optimal for medical images.

6. Applications and Transfer Learning

EPCA modules are general plug-and-play components:

• In general vision, they strengthen multi-scale contextual representation in image classification, object detection, and instance segmentation.
• For medical imaging, explicit modeling of global-local context and efficient channel fusion enables data-efficient transfer, especially under regimes where only the adapter (EPCA) is finetuned atop frozen pretrained backbones (Zhang et al., 2023).

This adapter scenario achieves nearly full finetuning performance with orders-of-magnitude fewer parameters, highlighting EPCA’s suitability for parameter-efficient transfer learning.

7. Design Insights and Comparative Evaluation

Key findings from ablation and comparative experiments include:

• Multi-scale context fusion, carefully scheduled group sizes, and concatenation-based branch fusion confer consistent performance gains.
• Softmax normalization across scales in EPSA and channel-wise linear fusion in EPCA allow adaptive selection of spatial context per channel.
• Compared to SE, CBAM, SPA, and ECA, both EPSA and pooling-fusion EPCA yield accuracy improvements of 0.7–2.2% on complex benchmarks with minimal cost.
• Theoretical and empirical results demonstrate that squeezing channels and leveraging grouped convolutions constrain parameter/FLOP growth linearly, unlike dense, fully connected recalibration layers found in alternative modules.

In summary, Efficient Pyramid Channel Attention (EPCA) modules implement scalable, multi-scale, low-cost mechanisms for context-aware channel recalibration, demonstrating robust accuracy improvements across both natural and medical image analysis tasks, and enabling high-impact, parameter-efficient transfer learning scenarios (Zhang et al., 2021, Zhang et al., 2023).