Image-Retrieval

Image retrieval program made in Tensorflow supporting the pretrained networks VGG16, VGG19, InceptionV3 and InceptionV4 and own trained Convolutional Autoencoder that you can train with this tool.

Requirements

• Python 3.*
• Tensorflow
• Pillow
• tqdm
• Pretrained VGG16, VGG19, InceptionV3 or InceptionV4 network or own trained Convolutional Autoencoder.

Usage

Firstly put your images in the images folder.

Embedding images and saving them

Get embedding from trained Convolutional autoencoder To train a Convolutional autoencoder to vectorize images do this command:

python3 autoencoder_training.py

You can get a look at the hyperparameters using.

python3 autoencoder_training.py --help

The same principles follow in all the other scripts.

Embedding with autoencoder Just do this command.

python3 vectorize_autoencoder.py

Get embedding from pretrained models Just do this command.

python3 vectorize_pretrained.py --model_path=<model_path> --model_type=<model_type> --layer_to_extract=<layer_to_extract>

What does these arguments mean?

model_path: Path to pretrained model. e.g ./inception_v4.ckpt

model_type: Type of model, either VGG16, VGG19, InceptionV3 or InceptionV4. e.g InceptionV4

layer_to_extract: Which layer to take vector from. e.g Mixed_7a

This command will save the vectors in a file in the vectors folder and will print out the path to the vectors for later use or evaluation at the end of the program.

Evaluating To evaluate your vectors you can do this command.

python3 evaluation.py --vectors_path=<vectors_path> --image_path=<image_path>

What does these arguments mean?

vectors_path: Where vectors are saved. e.g vectors/vectors_1

image_path: Image to evaluate on, i.e the image to check nearest neighbour on. e.g img.jpg

Todos

• Dimensionality reduction using PCA.
• More ways of doing NN search.
• Clean the code!
• Adversarial loss on autoencoder.

Other

Made by Oliver Edholm, 14 years old.