One-shot logo detection with keras

Logo detection in images has many applications, such as in advertisement and social media marketing. Existing methods usually require huge amount of training data for every logo class, since they are based on standard machine learning paradigms. They learn to recognize logos from examples of each logo class. Obviously, if you didn't train a model with your logo, you cannot search for it.

For practical application, it is more useful to search for logos in images only providing a query logo (i.e. an example of the logo), a machine learning paradigm known as "one-shot learning". This is what this repo is about.

This project is an implementation of the paper "A Deep One-Shot Network for Query-based Logo Retrieval" (Bhunia et al. 2019, https://arxiv.org/pdf/1811.01395). The algorithm searches for logos within a given target image, and predicts its presence and location (with a segmentation mask).

The neural network model is composed by a "conditional" branch and a "segmentation" branch. The conditional branch gives a latent representation of the query logo. This representation is then combined with feature maps of the segmentation branch at multiple scales, making the model scale-invariant.

Illustration of query-based logo detection problem (from Bhunia et al. 2019)

Depencencies

Install the libraries using:

pip install -r requirements.txt 

Weights

The graph and trained weights can be found at:

https://drive.google.com/drive/folders/1ojz6i0dsEEOzJ3qaAeutqD64x5dsxVvF?usp=sharing

Results

Here a screenshots of the prediction for a logo of the same class:

And here the prediction for a different class. The predicted mask is zero.

Train

To train a model run:

python main.py -c configs/config.yml --train

If you set "weights_initialization" in config.yml you can use a pretrained model to inizialize the weights, so you can stop the training and restore latere if needed.

During training the best and last snapshots are saved if you set those options in "callbacks" in config.yml.

Inference

To predict on a single image you use this command:

python main.py -c configs/config.yml --predict --query query_path --image filename_path --output result.jpg

You can use predict.sh script to launch a quick test of inference

To do

• [x]

References

[1] A Deep One-Shot Network for Query-based Logo Retrieval

[2] One-Shot Learning introduction (from Medium)