Image Recognizer (TensorFlow, CNN)

This is Image Recognizer for own images by using TensorFlow & CNN model (CIFR-10 tutorial). You can gather images, convert them to fit learning model's input. and then train. Finally, you launch web application on localhost, the application classify images you upload into the application.

Processes

1. Gather images you want to recognize from TensorFlow and put into /data/raw
2. Process the images to CIFAR-10 dataset (data/input/images.bin)
3. Train CIFAR-10's model using by your images.
4. Classify from New data.

Versions

Mac OS Sierra (10.12)
Ruby 2.2.2
Python 3.4.5

Usage

1. [Optional] Download images you want to recognize.

You don't need download any images because the sample images already exist in data/raw. If you need original images and recognize them, execute below commands.

# install chrome binary and move to PATH directory
# Download url: http://chromedriver.storage.googleapis.com/index.html

# gem installs
$ bundle install

# Scrape google images by using `keyword`
# num is less than 400 because of bugs
$ ./scrape/scrape.rb -k <keyword> -n <num>

# and then remove the images not matched with keyword.

2. Process your images to fit TensorFlow CNN model dataset (CIFAR-10)

Process your images to CIFAR-10 binary

# Options
#   -s src directory
#   -o output directory
#   (Be carefull directory path starts with '.' and doesn't end with '/')
$ ./image_processor/processor.rb -s ./data/raw -o ./data/input

3. Train images by using TensorFlow CNN model.

Run training (training will take 1 hours. depends on machine power) Be careful to re-run this command. because checkpoint file(trained data) is removed and new one is created.

$ python recognizer/train.py
Filling queue with 40 CIFAR images before starting to train. This will take a few minutes.
2016-10-07 11:02:45.512924: step 0, loss = 4.77 (78.8 examples/sec; 0.813 sec/batch)
2016-10-07 11:02:55.742913: step 10, loss = 4.72 (87.8 examples/sec; 0.729 sec/batch)
2016-10-07 11:03:03.609812: step 20, loss = 4.70 (53.7 examples/sec; 1.191 sec/batch)
2016-10-07 11:03:10.959503: step 30, loss = 4.64 (85.9 examples/sec; 0.745 sec/batch)
2016-10-07 11:03:18.221603: step 40, loss = 4.58 (94.1 examples/sec; 0.680 sec/batch)
...

Visualize results by TensorBoard.

$ tensorboard --logdir=/tmp/my_model_train

4. Classify images as web application.

Measure accuracy by using test data.

$ cd recognizer
$ python model_eval.py

# if 2 classifications (cat or dog)
2016-10-07 13:30:44.203764: precision @ 1 = 0.802

# if 3 classification (cat, dog or bird)
2016-10-07 15:00:27.723453: precision @ 1 = 0.656

nipe0324/tf_image_recognizer