Using machine learning to identify old logos

Some tasks are easy to automate, such as checking when your site was last modified. Or how old a file is. But how do you check whether the image contains an old logo?

In the past this is a task I would’ve done manually, to ensure every old image is found. But with machine learning it’s possible to train a neural network to identify two different scenarios and then have it detect which image it is closest to.

To accomplish this I started off using a Tensorflow code lab as a starting point:
https://codelabs.developers.google.com/codelabs/tensorflow-for-poets/#0

1. Install python dependencies. Create a file /requirements.txt and add the list of libraries required:

absl-py==0.7.0
astor==0.7.1
backports.weakref==1.0.post1
enum34==1.1.6
funcsigs==1.0.2
futures==3.2.0
gast==0.2.2
grpcio==1.19.0
h5py==2.9.0
Keras-Applications==1.0.7
Keras-Preprocessing==1.0.9
Markdown==3.0.1
mock==2.0.0
numpy==1.16.2
pbr==5.1.3
protobuf==3.7.0
six==1.12.0
tensorboard==1.13.0
tensorflow==1.13.1
tensorflow-estimator==1.13.0
termcolor==1.1.0
Werkzeug==0.14.1

2. Create a python virtual environment and install the dependencies using:

virtualenv env
source env/bin/activate
pip install -r requirements.txt

3. Prepare your data. This is arguably the biggest task you’ll have, to get enough source files to give the script access to a variety of examples. Create two folders containing old and new examples of the logo.

/src/images/new/
/src/images/old/

4. Next create your training script. This will navigate to each image subdirectory and use the name as a classifier group, in this example two groups ‘old’ and ‘new’. I have prepared a script you can use at /src/retrain.py

5. Run the training script using the following command to train the image model:

python -m src.retrain \
    --bottleneck_dir=dist/bottlenecks \
    --how_many_training_steps=500 \
    --model_dir=dist/models \
    --summaries_dir=dist/summaries \
    --output_graph=dist/retrained_graph.pb \
    --output_labels=dist/retrained_labels.txt \
    --architecture="mobilenet_0.50_224" \
    --image_dir=src/images \
    --learning_rate=0.005

6. Create the classifier script. You can use this script at /src/classify.py

7.Prepare some test images which are not part of the training data set. These will allow you to test the confidence scores of the results on unseen examples:

/src/images_test

8. Run the classifier script using the command:

python -m src.classify \
    --graph=dist/retrained_graph.pb  \
    --labels=dist/retrained_labels.txt \
    --image=src/images_test/o1.jpg

Check your results. In my case I get the following:

Our image classifier has a high confidence detecting our examples.

Hope that helps you get started with image classifiers!

Check out the full source code and example at:
https://github.com/kmturley/tensorflow-detect-logo