Skip to content

Wavelet convolutional neural network combines a multi-resolution analysis and convolutional neural network into a single model to achieve computer vision task.

Notifications You must be signed in to change notification settings

happymondaynkanta/Wavelet-Convolutional-Neural-Network-Using-Keras

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

17 Commits
 
 
 
 

Repository files navigation

Wavelet-Convolutional-Neural-Network

Wavelet convolutional neural network combines a multiresolution analysis and convolutional neural network into a single model to achieve computer vision task.

Model Architecture of wavelet convolutional neural network

wavelet-cnn

Description wavelet convolutional neural network connectivity

Overview of wavelet CNN with 4-level decomposition of the input image. Wavelet CNN processes the input image through convolution layers with 3 × 3 kernels and 1 × 1 padding. The number after Conv denotes the number of channels of the output. 3 × 3 convolutional kernels with the stride of 2 and 1 × 1 padding are used to reduce the size of feature maps. Additionally, the input image is decomposed through multiresolution analysis and the decomposed images are concatenated channel-wise. The projection shortcuts are done by 1 × 1 convolutions. The output of convolution layers is vectorized by global average pooling followed by a fully connected layer (fc). The size of the output is equal to the number of classes included in the input dataset. Click here to read the paper

Multi-Resolution Analysis of Discrete Wavelet Transform Decomposition

For 2-dimensional image input, Multi-Resolution Analysis of Discrete Wavelet Transform Decomposition is applied as a technique to decomposed the image into approximate and detail coefficient of varying frequencies and scales through low- and high-pass filters with the scale factor of 2. The approximat coefficient is the low frequency component while the detail coefficient is the high frequency component.

A

Approximate image (LL)

H

Horizontal image (HL)

V

Vertical image (LH)

D

Diagonal image (HH)

At each level of decomposition, sub-bands (LL, HL, LH, HH) of images are generated which are fed as input to the neural network. For subsequent decomposition levels, the approximate coefficient which is the low-frequency component is further decomposed to generate another sub-bands of images.

LEVELS

Input flow into the neural network

sub-bands

Two level wavelet decomposition

About

Wavelet convolutional neural network combines a multi-resolution analysis and convolutional neural network into a single model to achieve computer vision task.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published