A model trained for watermark detection using Tensorflow.
The code and description can be found in watermark-detector.ipynb.
The dataset is expected to be in the same directory as the notebook. But you can change base_dir variable in the notebook to point to the arbitrary dataset parent directory.
Another version of code using pre-trained ResNet50 on "imagenet" dataset as the base model for transfer learning is also available in watermark-detector-transfer-learning.ipynb.
The dataset contains images from Digikala, an Iranian online shop, some including various kinds of watermarks. The data is labeled with positive or negative, indicating whether the image has or has not watermark respectevly.
The dataset structure is as follows:
dataset
/ train
/ positive
/ negative
/ test
Data can be accessed from here.
There are 8582 samples for training and 1103 unlabeled samples for testing.
You can submit your testing prediction result in Quera (the problem statement is in Persian) and get the accuracy of the model on the test data.