Classifying Dog or Cat using Bag Of Words Concept

This is a proof of concept to apply Bag of Words model in Natural Language Processing into classifying images, in this case to classify between dog or cat.

Dataset

The dataset is obtained from Kaggle: Dog vs Cat containing 25000 images of a dogs and cats. Each image has the label as the part of the filename.

Building Codebook

A codebook (or dictionary) can be built from all the images in the dataset with some parameters below:

• The full path to the root directory containing all the images
• The output file full path
• Descriptors algorithm to obtain descriptors of all images (either SIFT or KAZE)
• The number of vocabularies

Format:

python codebook.py -i [images_dir] -o [output_file] -a [sift|kaze] -s [vocab_size] --verbose(optional)

Example:

python codebook.py -i /home/agumbira/dev/data/dog_cat_kaggle/train -o /home/agumbira/dev/python/BOWImageClassifier/model/dog_cat_kaggle/codebook_kaze_200.pkl -a kaze -s 200 --verbose

Finding the best network model

The best network model is found by running hyperparameter_optimization .py. The search scope is defined in the space variable. 1000 network models with various parameters are run and the one with the highest ROC AUC is selected. This model then used for training.

Training

To train the network model (found by hyperparameter optimization), run python training.py