# Wavelet Convolutional Neural Networks

**Authors:** Shin Fujieda, Kohei Takayama, Toshiya Hachisuka  
**arXiv:** [1805.08620](https://arxiv.org/abs/1805.08620)  
**Date:** 2018-05-20  
**Summary page:** https://www.summarizepaper.com/en/paper/1805.08620v1/

## Key points

- Spatial and spectral approaches are widely used in image processing for tasks like classification and recognition
- Integration of spectral approaches into CNNs is gaining interest for leveraging their distinct characteristics
- The study "Wavelet Convolutional Neural Networks" proposes a novel architecture combining multiresolution analysis with CNNs to create wavelet CNNs
- Wavelet CNNs incorporate wavelet transform to supplement missing parts of multiresolution analysis, allowing utilization of spectral information lost in traditional CNNs
- Experimental evaluation shows superior accuracy and fewer parameters required compared to traditional CNN architectures
- Wavelet CNNs have the potential to enhance performance in image processing tasks by integrating spatial and spectral information effectively

## Summary

In the field of image processing, spatial and spectral approaches are widely used for tasks such as image classification and object recognition. These methods have shown significant performance improvements in challenging tasks. However, there is growing interest in integrating spectral approaches into convolutional neural networks (CNNs) to leverage their distinct characteristics. In this study titled "Wavelet Convolutional Neural Networks," authors Shin Fujieda, Kohei Takayama, and Toshiya Hachisuka propose a novel CNN architecture that combines multiresolution analysis with CNNs to create wavelet CNNs. The key insight behind this approach is that a CNN can be seen as a limited form of multiresolution analysis. By incorporating wavelet transform to supplement missing parts of the multiresolution analysis, the authors introduce additional components into the architecture that allow for the utilization of spectral information typically lost in conventional CNNs. The experimental evaluation of wavelet CNNs on texture classification and image annotation tasks demonstrates superior accuracy compared to existing models while also requiring significantly fewer parameters than traditional CNN architectures. This highlights the potential of wavelet CNNs to enhance performance in image processing tasks by effectively integrating spatial and spectral information. Overall, this research contributes to advancing the field of image processing by introducing a novel approach that bridges spatial and spectral methodologies through the innovative combination of multiresolution analysis and convolutional neural networks.