WaveletCNN for Texture Classification

This is a Caffe implementation of a paper, Wavelet Convolutional Neural Networks for Texture Classification (arXiv, 2017).

Requirements

• Python 3+
• Caffe

This code was tested with NVIDIA GeForce GTX 1080 on Windows 10.

When you set up the environment of Caffe on Windows, we recommend you to use this version of Caffe, instead of the windows branch of the official Caffe.

Pre-trained Model

Please go to models directory and follow the instructions.

Usage

Train with your own dataset

python run_waveletcnn.py --phase train --gpu 0 --dataset path/to/dataset_lmdb

To run this code, you have to prepare your dataset as LMDB format. (Otherwise, you have to rewrite models/WaveletCNN_4level.prototxt)

And you might need to rewrite some settings in models/solver_WaveletCNN_4level.prototxt, such as test_iter, base_lr and max_iter.

Test with an image

python run_waveletcnn.py --phase test --gpu 0 --initmodel models/ImageNet_waveletCNN_level4.caffemodel --target_image path/to/image

Using the pre-trained model with ImageNet 2012 dataset, please download it following the instructions in models directory.

When you use your own trained model, path to the model is used instead as the argument of --initmodel. Additionally, you need to rewrite the path to the file, which contains label names, in run_waveletcnn.py (l.68).

Citation

If you find this code useful for your research, please cite our paper:

@article{Fujieda2017,
  author    = {Shin Fujieda and Kohei Takayama and Toshiya Hachisuka},
  title     = {Wavelet Convolutional Neural Networks for Texture Classification},
  journal   = {arXiv:1707.07394 [cs.CV]},
  year      = {2017},
}