WaveCNet

WaveCNet enhances noise robustness in image classification by integrating discrete wavelet transform (DWT) and inverse DWT (IDWT) layers into convolutional neural network (CNN) architectures to separate low-frequency signal from high-frequency noise.

Key Features:

• DWT/IDWT down-sampling: Replaces max-pooling, strided convolution, and average pooling with discrete wavelet transform and inverse transform layers.
• Wavelet support: Compatible with orthogonal and biorthogonal discrete wavelets including Haar, Daubechies, and Cohen.
• Architecture integration: Incorporated into VGG, ResNets, and DenseNet to form Wavelet Integrated CNNs (WaveCNets).
• Frequency decomposition: Decomposes feature maps into low-frequency and high-frequency components, retaining low-frequency components and discarding high-frequency components.
• Aliasing suppression: Reduces aliasing effects by separating low-frequency information from high-frequency noise.
• Robustness gains: Demonstrates improved accuracy on ImageNet and increased robustness on ImageNet-C and against six adversarial attacks.
• Object detection backbone: When used as a backbone in Faster R-CNN and RetinaNet, yields improved performance on the COCO detection dataset.

Scientific Applications:

• Image classification: Enhances classification accuracy on ImageNet when integrated into CNN backbones.
• Noise and adversarial robustness testing: Improves performance on ImageNet-C and under six adversarial attack scenarios.
• Object detection: Serves as a backbone for Faster R-CNN and RetinaNet to improve detection results on the COCO dataset.

Methodology:

WaveCNet replaces traditional down-sampling (max-pooling, strided convolution, average pooling) with DWT/IDWT layers that decompose feature maps into low- and high-frequency components, transmit low-frequency components through subsequent layers, discard high-frequency components, and support orthogonal and biorthogonal wavelets (Haar, Daubechies, Cohen).