/low_res_logos_gan

DC GAN for low resolution logo generation

Primary LanguageJupyter Notebook

DC GAN for low resolution logo generation

Attached notebook contains an implementation (tensorflow 2.2.0) of vanilla/least squares DC GAN for generating low resolution logos.

Generated examples:

Generated Logos

Training data contains 12180 images parsed from the web (for downloading the full data parser from this repository was used). To simplify generation process for vanilla GAN, all logos containing text were filtered out using OpenCV and EAST text detector. I used this pre-trained model: https://raw.githubusercontent.com/oyyd/frozen_east_text_detection.pb/master/frozen_east_text_detection.pb. This notebook will walk you through text detection. For the purpose of GAN training data was augumented with random spins, flips and hue adjustments.

Data augumentation:

Data augumentation 1

Data augumentation 2

Training dynamics:

Epoch-1
Epoch-1

Epoch-50
Epoch-50

Epoch-130
Epoch-130

Epoch-200
Epoch-200

Epoch-300
Epoch-300

Epoch-500
Epoch-500