Deep learning batik classification

This repo contains programs that are used in experiment of paper: Batik Classification using Deep Convolutional Network Transfer Learning

Indonesian Batik classification using VG16 convolutional neural network (CNN) as feature extractor + softmax as classifier.

Our dataset consists of 5 batik classes where each images will belong to exactly one class:

1. Parang: parang (traditional blade) motifs
2. Lereng: also blade-like pattern but less pointy than Parang
3. Ceplok: repetition of geometrical motif shapes (eg. rectangle, square, ovals, stars)
4. Kawung: kawung fruits motif
5. Nitik: flowers-like motifs

Requirements

To run the experiments, you would need:

• CUDA device with global RAM >= 4 GB (tested with GTX 980)
• Python 2.7.x
• Virtualenv (optional. For isolated environment)

The programs can also run on CPU. You just need to use tensorflow instead of tensorflow-gpu. Further info https://www.tensorflow.org/install/

The programs expect batik image files (jpg/png) to be grouped by their classes. In this experiment, below directory structures are used:

train_data_dir/
	Ceplok/*.jpg
	Kawung/*.jpg
	Lereng/*.jpg
	Nitik/*.jpg
	Parang/*.jpg

test_data_dir/
	Ceplok/*.jpg
	Kawung/*.jpg
	Lereng/*.jpg
	Nitik/*.jpg
	Parang/*.jpg

Installation

Install dependencies: pip install -r requirements.txt

Experiments

VGG16 + Softmax NN

Following commands are executing these steps:

1. Convert images dataset to vector h5 format:
2. Extract features from dataset (still in h5 format)
3. Train & evaluate variations of NN (vary by number of hidden layers and activation function) with cross validation

python preprocess.py train_data_dir --vector_file train.h5
python preprocess.py test_data_dir --vector_file test.h5
python extractor.py train.h5 --features_file train.features.h5
python extractor.py test.h5 --features_file test.features.h5
python evaluate.py train.features.h5 test.features.h5

For more options run each file with -h option. eg: python preprocess.py -h

Process time is around 40 minutes with GTX 980 4 GB

VGG16 + scikit-learn classifiers

Follow the same steps as VGG16 + Softmax NN, but use evaluate_sklearn.py instead of evaluate.py:

python preprocess.py train_data_dir --vector_file train.h5
python preprocess.py test_data_dir --vector_file test.h5
python extractor.py train.h5 --features_file train.features.h5
python extractor.py test.h5 --features_file test.features.h5
python evaluate_sklearn.py train.features.h5 test.features.h5

Classification using SIFT Bag of Words + SVM

Following command is executing these steps:

1. Extract SIFT descriptors from images
2. Cluster descriptors to build vocabulary using K-means
3. Extract bag of words features from images using vocabulary
4. Train & evaluate 6 classifiers with cross validation

python siftbow.py train_data_dir test_data_dir

For more options run each file with -h option. eg: python siftbow.py -h

Process time is around 52 minutes with Intel Core i7 5500U 8 GB RAM

Classification using SURF Bag of Words + SVM

Following command is executing these steps:

1. Extract SURF descriptors from images
2. Cluster descriptors to build vocabulary using K-means
3. Extract bag of words features from images using vocabulary
4. Train & evaluate 6 classifiers with cross validation

python surfbow.py train_data_dir test_data_dir

For more options run each file with -h option. eg: python surfbow.py -h

Process time is around 45 minutes with Intel Core i7 5500U 8 GB RAM

Citations

To cite our research or dataset, please use this citation:

@article{gultom2018batik,
  title={Batik Classification using Deep Convolutional Network Transfer Learning},
  author={Gultom, Yohanes and Arymurthy, Aniati Murni and Masikome, Rian Josua},
  journal={Jurnal Ilmu Komputer dan Informasi},
  volume={11},
  number={2},
  pages={59--66},
  year={2018}
}