[Paper Review] Wavelet Convolutional Neural Networks for Texture Classification

Shin Fujieda, Kohei Takayama+1 more|arXiv (Cornell University)|Jul 24, 2017

Image Retrieval and Classification TechniquesComputer Science27 references100 citations

TL;DR

Introduces wavelet CNNs that integrate multiresolution spectral analysis into CNNs to improve texture classification with fewer parameters.

ABSTRACT

Texture classification is an important and challenging problem in many image processing applications. While convolutional neural networks (CNNs) achieved significant successes for image classification, texture classification remains a difficult problem since textures usually do not contain enough information regarding the shape of object. In image processing, texture classification has been traditionally studied well with spectral analyses which exploit repeated structures in many textures. Since CNNs process images as-is in the spatial domain whereas spectral analyses process images in the frequency domain, these models have different characteristics in terms of performance. We propose a novel CNN architecture, wavelet CNNs, which integrates a spectral analysis into CNNs. Our insight is that the pooling layer and the convolution layer can be viewed as a limited form of a spectral analysis. Based on this insight, we generalize both layers to perform a spectral analysis with wavelet transform. Wavelet CNNs allow us to utilize spectral information which is lost in conventional CNNs but useful in texture classification. The experiments demonstrate that our model achieves better accuracy in texture classification than existing models. We also show that our model has significantly fewer parameters than CNNs, making our model easier to train with less memory.

Motivation & Objective

• Motivate improved texture classification by incorporating spectral analysis into CNNs.
• Generalize pooling and convolution to perform wavelet-based multiresolution analysis.
• Demonstrate accuracy and parameter efficiency on standard texture datasets.
• Compare with AlexNet, T-CNN, and spectral methods to show benefits of the approach.

Proposed method

• Reformulate convolution and pooling as generalized filtering and downsampling.
• Incorporate Haar wavelet multiresolution analysis into the network, using low/high frequency components.
• Use a VGG-19-like architecture with 3x3 convolutions, 1x1 padding, and stride-based downsampling.
• Insert an energy layer before fully connected layers to enhance texture features.
• Train from scratch and with ImageNet pre-training to compare performance.
• Implement in Caffe and train on 224x224 inputs with data augmentation and batch normalization.

Experimental results

Research questions

• RQ1 Can wavelet-based multiresolution analysis within a CNN improve texture classification accuracy compared to traditional CNNs?
• RQ2 Does incorporating high-frequency components alongside low-frequency components preserve information and improve robustness to texture variation?
• RQ3 How does the wavelet CNN compare to existing spectral and CNN-based texture methods in terms of accuracy and parameter efficiency?
• RQ4 What is the impact of decomposition level on texture classification performance?

Key findings

• Wavelet CNNs outperform AlexNet and T-CNN on texture datasets when trained from scratch across multiple decomposition levels.
• The 4-level decomposition often yields the best balance of accuracy and parameters.
• With ImageNet pre-training, wavelet CNNs achieve top performance on kth-tips2-b and competitive results on DTD, with far fewer parameters than FV-CNN.
• Wavelet CNNs use substantially fewer trainable parameters than competing models (e.g., under 90 MB in some configurations).
• Four-level decomposition provided strong performance while five-level offered diminishing returns due to increased parameters.