[Paper Review] Wavelet-Attention CNN for Image Classification

Xiangyu Zhao|arXiv (Cornell University)|Jan 23, 2022

Advanced Neural Network Applications4 citations

TL;DR

This paper proposes Wavelet-Attention CNN (WA-CNN), a novel convolutional neural network that enhances image classification by applying a Wavelet-Attention (WA) block to selectively focus on high-frequency details while preserving low-frequency structural features via Discrete Wavelet Transform (DWT). The WA block improves feature learning by filtering noise and emphasizing relevant high-frequency information, achieving 1.26% and 1.54% Top-1 accuracy gains on CIFAR-10 and CIFAR-100 using MobileNetV2 backbones.

ABSTRACT

The feature learning methods based on convolutional neural network (CNN) have successfully produced tremendous achievements in image classification tasks. However, the inherent noise and some other factors may weaken the effectiveness of the convolutional feature statistics. In this paper, we investigate Discrete Wavelet Transform (DWT) in the frequency domain and design a new Wavelet-Attention (WA) block to only implement attention in the high-frequency domain. Based on this, we propose a Wavelet-Attention convolutional neural network (WA-CNN) for image classification. Specifically, WA-CNN decomposes the feature maps into low-frequency and high-frequency components for storing the structures of the basic objects, as well as the detailed information and noise, respectively. Then, the WA block is leveraged to capture the detailed information in the high-frequency domain with different attention factors but reserves the basic object structures in the low-frequency domain. Experimental results on CIFAR-10 and CIFAR-100 datasets show that our proposed WA-CNN achieves significant improvements in classification accuracy compared to other related networks. Specifically, based on MobileNetV2 backbones, WA-CNN achieves 1.26% Top-1 accuracy improvement on the CIFAR-10 benchmark and 1.54% Top-1 accuracy improvement on the CIFAR-100 benchmark.

Motivation & Objective

• To address the degradation of feature representation in CNNs due to noise and redundant high-frequency components in feature maps.
• To improve feature learning in image classification by selectively attending to high-frequency details while preserving low-frequency structural information.
• To design a frequency-aware attention mechanism that operates only in the high-frequency domain, minimizing interference with main feature structures.
• To validate the effectiveness of the proposed Wavelet-Attention block across multiple backbone architectures on standard benchmarks.
• To demonstrate that frequency-domain attention via DWT can enhance model generalization and reduce overfitting in small-scale datasets.

Proposed method

• The method decomposes feature maps into low-frequency and high-frequency components using Discrete Wavelet Transform (DWT), separating structural content from detailed and noisy information.
• A Wavelet-Attention (WA) block is introduced to apply learnable attention weights exclusively to the high-frequency components, enhancing relevant details while leaving low-frequency components untouched.
• The WA block uses a learnable attention mechanism that adaptively recalibrates high-frequency features based on their importance, improving representation without altering the main structural features.
• The WA-CNN integrates the WA block into existing CNN backbones such as VGG, ResNet, and MobileNetV2, enabling end-to-end training with standard optimization protocols.
• The model is trained using standard data augmentation (random crop, flip) and a cosine decay learning rate schedule on CIFAR-10 and CIFAR-100 datasets.
• Feature reconstruction is performed via inverse DWT after attention, ensuring compatibility with standard CNN layers and training pipelines.

Experimental results

Research questions

• RQ1 Can frequency-domain decomposition via DWT improve the robustness and accuracy of CNN-based image classification by isolating noise and detail in high-frequency components?
• RQ2 Does applying attention only in the high-frequency domain enhance feature representation without degrading the main structural information in low-frequency components?
• RQ3 How does the WA-CNN compare to standard attention mechanisms (e.g., SE, CBAM) in terms of accuracy and generalization on small-scale datasets?
• RQ4 Can the WA block be effectively integrated into various CNN architectures without architectural overhaul?
• RQ5 Does the WA-CNN reduce overfitting on small datasets by filtering out high-frequency noise while preserving meaningful edge and texture details?

Key findings

• WA-CNN achieves a 1.26% Top-1 accuracy improvement over MobileNetV2 on the CIFAR-10 benchmark, with a 1.54% improvement on CIFAR-100.
• On VGG16bn, the WA block achieves the highest Top-1 accuracy of 93.89% on CIFAR-10 and 73.66% on CIFAR-100, outperforming all baseline and attention-based models.
• The WA block reduces training loss by approximately 0.2 compared to the baseline VGG16bn during the first 80 epochs, indicating faster and more stable convergence.
• The WA block consistently improves performance across all tested backbones (VGG, ResNet, MobileNetV2), with the most significant gains on lightweight models like MobileNetV2.
• The WA block achieves competitive results compared to other attention mechanisms (SE, CBAM, ECA), outperforming them in most settings, especially on larger networks.
• Visualizations show that WA-CNN improves recognition of similar or fine-grained images, indicating enhanced discriminative capability.