Source-linked AI summary
EPSANet: An Efficient Pyramid Squeeze Attention Block on Convolutional Neural Network
Hu Zhang, Keke Zu, Jian Lu, Yuru Zou, Deyu Meng
TL;DR
Deep convolutional networks can benefit from attention, but efficient multi-scale spatial and channel modeling remains challenging. This paper introduces PSA, builds EPSA blocks by replacing ResNet bottleneck convolutions, and stacks them into EPSANet. The resulting backbone is reported to outperform most state-of-the-art channel attention methods across image classification, object detection, and instance segmentation.
Problem
Existing attention methods motivate efficient modeling of multi-scale spatial information and channel dependencies without imposing heavy computational complexity.
Method
The paper proposes PSA, integrates it into ResNet bottlenecks as EPSA blocks, and stacks those blocks to form the EPSANet backbone.
Results
EPSANet achieves promising results across image classification, object detection, and instance segmentation on ImageNet and COCO, outperforming most state-of-the-art channel attention methods.
Takeaways & Limitations
EPSANet provides a flexible and scalable backbone for richer multi-scale representations across multiple computer vision tasks.
Abstract
from arXiv · showhide
Recently, it has been demonstrated that the performance of a deep convolutional neural network can be effectively improved by embedding an attention module into it. In this work, a novel lightweight and effective attention method named Pyramid Squeeze Attention (PSA) module is proposed. By replacing the 3x3 convolution with the PSA module in the bottleneck blocks of the ResNet, a novel representational block named Efficient Pyramid Squeeze Attention (EPSA) is obtained. The EPSA block can be easily added as a plug-and-play component into a well-established backbone network, and significant improvements on model performance can be achieved. Hence, a simple and efficient backbone architecture named EPSANet is developed in this work by stacking these ResNet-style EPSA blocks. Correspondingly, a stronger multi-scale representation ability can be offered by the proposed EPSANet for various computer vision tasks including but not limited to, image classification, object detection, instance segmentation, etc. Without bells and whistles, the performance of the proposed EPSANet outperforms most of the state-of-the-art channel attention methods. As compared to the SENet-50, the Top-1 accuracy is improved by 1.93% on ImageNet dataset, a larger margin of +2.7 box AP for object detection and an improvement of +1.7 mask AP for instance segmentation by using the Mask-RCNN on MS-COCO dataset are obtained. Our source code is available at:https://github.com/murufeng/EPSANet.
1 Introduction
The paper proposes lightweight attention modules and an EPSANet backbone to capture multi-scale spatial information and cross-dimension channel dependencies across vision tasks. EPSANet is reported to achieve promising results in image classification, object detection, and instance segmentation.
- Motivation: Attention methods improve computer vision models, but existing approaches face challenges efficiently capturing spatial information at different scales and controlling computational complexity.The introduction identifies these as important challenges motivating a low-cost attention module.
- Proposed approach: PSA processes input features at multiple scales using pyramid convolutions, channel squeezing, and Softmax-based channel reweighting.This design integrates multi-scale spatial context and establishes long-range channel dependency.
- Proposed approach: EPSA extracts multi-scale spatial information at a granular level while developing long-range channel dependency.The block is described as flexible and scalable for diverse network architectures and computer vision tasks.
- Proposed approach: EPSANet learns richer multi-scale feature representations and adaptively recalibrates cross-dimension channel-wise attention weights.The architecture is formed as a backbone using the proposed EPSA design.
- Results: EPSANet achieves promising results across image classification, object detection, and instance segmentation on ImageNet and COCO.The reported evaluation spans three major computer vision tasks and two benchmark datasets.
2 Related Work
The related work frames attention as selective feature weighting and emphasizes multi-scale feature representation as essential for visual recognition. Existing attention mechanisms address channel or spatial information, while CNNs naturally learn coarse-to-fine features through stacked convolutions.
- Attention mechanism: Attention mechanisms selectively weight informative channels or regions while suppressing less useful feature expressions.SE modulates channel scales, whereas CBAM enriches attention using pooled features and large kernels.
- Multi-scale Feature Representations: Multi-scale feature representation is essential for tasks including instance segmentation, object detection, and semantic segmentation.The related work motivates efficient operators that extract multi-scale features within convolutional networks.
- Multi-scale Feature Representations: CNNs naturally learn coarse-to-fine multi-scale features through stacked convolutional operators, making convolutional operator design important for representation quality.The passage presents operator design as a key direction for improving multi-scale representations.
3 Method
The method builds PSA to combine multi-scale spatial features with cross-channel attention, then forms EPSA blocks and the EPSANet backbone by replacing ResNet bottleneck 3x3 convolutions. Its components use parallel grouped convolutions, scale-specific attention, and concatenation to produce refined multi-scale representations.
- SPC Module: The SPC module extracts multi-scale spatial information through parallel branches with different convolutional kernels and grouped convolutions.Grouping is introduced to process larger kernel sizes without increasing computational cost; the input channels are equally squeezed across branches.
- PSA Module: SEWeight generates attention vectors for feature maps at different scales, which are concatenated into a multi-scale channel attention vector.The concatenated vector supports interaction between local and global channel attention while preserving the original channel attention information.
- PSA Module: Softmax re-calibrates multi-scale channel weights before they are multiplied with their corresponding feature maps and concatenated into the refined output.The concatenation operator is described as preserving feature representation more completely than summation.
- PSA Module: PSA processes feature maps at multiple scales and combines spatial information with cross-channel attention.The module uses scale-specific feature maps, attention weights, softmax-based re-calibration, and feature re-calibration.
- EPSA and EPSANet: EPSA replaces the 3x3 convolution in ResNet bottleneck blocks with PSA, and EPSANet stacks EPSA blocks as an efficient backbone.The EPSA block is intended to extract multi-scale spatial information at a finer level and develop long-range channel dependency.
4 Experiments
Experiments evaluate EPSANet on ImageNet and COCO across classification, detection, and instance segmentation, showing strong accuracy and efficiency compared with established attention networks.
- Image Classification on ImageNet: 2.29% higher Top-1 accuracy than ResNet-50 is achieved by EPSANet-50(Small), with 11.7% fewer parameters and 12.1% lower computational cost.The comparison is reported on ImageNet.
- Image Classification on ImageNet: EPSANet-50(Large) exceeds SENet, ECANet, and FcaNet by about 1.93%, 1.16%, and 0.12% in Top-1 accuracy, respectively.EPSANet-101(Large) also improves over SENet101 and CBAM with the same number of parameters.
- Object Detection on COCO: EPSANet models achieve the best reported object-detection performance, while EPSANet-50(Small) improves bounding box AP over SENet-50 by 1.5%, 1.3%, and 1.1% across three detectors.The same model reduces parameters to 87.5%, 88.3%, and 86.4% of SENet-50 across Faster R-CNN, Mask R-CNN, and RetinaNet, respectively.
- Instance Segmentation on COCO: EPSANet-50(Large) surpasses FcaNet by about 0.9% AP, 0.4% AP50, and 1.4% AP75 for Mask R-CNN instance segmentation.Experiments use the MS COCO dataset.
- Ablation Study: Increasing kernel sizes in parallel significantly increases parameter count, motivating group convolution to exploit multi-scale location information without increasing computational cost.The effect of group size is evaluated on ImageNet.
5 Conclusion
The paper proposes PSA, EPSA, and EPSANet as lightweight mechanisms for multi-scale attention and representation learning. Experiments support state-of-the-art performance across classification, detection, and instance segmentation, while future work will examine lighter CNN architectures.
- Conclusion: PSA is an effective, lightweight, plug-and-play attention module that extracts multi-scale spatial information and cross-dimensional channel features.The paper positions PSA as applicable across computer vision tasks.
- Conclusion: EPSA improves multi-scale representation at a granular level and develops long-range channel dependency.This block is the representational unit used in the proposed architecture.
- Conclusion: EPSANet integrates multi-scale contextual features and image-level categorical information, achieving state-of-the-art results across image classification, object detection, and instance segmentation.The conclusion compares it with conventional channel attention methods.
- Conclusion: Future work will investigate adding the PSA module to more lightweight CNN architectures.This defines the stated scope of the proposed future investigation.