WaveCNet: Wavelet Integrated CNNs to Suppress Aliasing Effect for Noise-Robust Image Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Convolutional Neural Networks (CNNs) are susceptible to noise, impacting image classification accuracy.
• Traditional down-sampling methods in CNNs can amplify noise and reduce robustness.
• Integrating wavelet transforms offers a potential solution for enhancing CNN noise resilience.

Purpose of the Study:

• To develop novel discrete wavelet transform (DWT) and inverse DWT (IDWT) layers for CNNs.
• To design and evaluate Wavelet integrated CNNs (WaveCNets) for improved noise and adversarial robustness.
• To assess the impact of WaveCNets on object detection performance.

Main Methods:

• Proposed general DWT/IDWT layers compatible with various wavelet types (Haar, Daubechies, Cohen).
• Integrated DWT into established CNN architectures (VGG, ResNets, DenseNet) to create WaveCNets.
• Replaced standard down-sampling with DWT to decompose feature maps into low and high-frequency components.

Main Results:

• WaveCNets demonstrated superior accuracy on ImageNet compared to standard CNNs.
• Significant improvements in noise-robustness and adversarial robustness were observed on ImageNet-C and adversarial attack datasets.
• Object detection models (Faster R-CNN, RetinaNet) using WaveCNets as backbones showed consistent performance gains on the COCO dataset.

Conclusions:

• The proposed DWT/IDWT layers effectively enhance CNNs' noise and adversarial robustness.
• WaveCNets offer a promising approach for developing more resilient deep learning models for image analysis.
• Separating low and high-frequency information via DWT is key to mitigating aliasing effects and improving feature representation.