Scalable Logo Recognition Using Deep Neural Networks

This paper explores the problem of open-set logo classification using both the few-shot and metric learning approaches. Three approaches are compared: a simple classifier, a metric learning classifier with A-Sfotmax loss function, and a Neural Statistician - generative few-shot model. For open-set classification, an Extreme Vector Machine is used that evaluates the distribution of hidden representations for class images. The model can determine the independent probabilities of image classes that the model has already seen during training. These probabilities allow us not only to evaluate the model’s confidence in class selection but also to identify images whose classes the model has not seen during training. To evaluate the ability of models to detect new classes, we use 202 classes with the lowest frequency in the QMUL-OpenLogo dataset. The best model for this task was the metric learning model, which revealed 89% of images of new classes with a 0.01 probability of being assigned to a class. To evaluate the ability of images in searching for similar images on new data, the test part of the FlickLogos-32 dataset is used. The metric learning model (with 0.77 mean precision at 1) also showed the best results here.