A brand logo detection system using region-based convolutional neural networks.
Implentation of Region-based Convolutional Neural Networks in Tensorflow, to detect and classify brand logos.
Note: Since this is my first version model, there are some remaining issues.
Example detection results are below.
Here are some failure cases.
python gen_bg_class.py: Generate train_annot_with_bg_class.txt file.python crop_and_aug.py: Crop brand logo images from the flickr27_logos_dataset and apply data augmentation method. Finally the dataset consists of 140137 images.python gen_train_valid_test.py: Generate(Split) train/valid/test set from the dataset.python train_deep_logo_cnn.py: Train the convolutional neural networks and save the trained model to disk.python test_deep_logo_cnn.py: Test the trained model (for Classification).
python detect_logo.py: Test the trained model (for Detection)
The network is based on this blog post. Same network is applied to this brand logo recognition task because a brand logo is similar to a number plate which consists of a number of digits and letters.
MIT