Scalable Logo Detection and Recognition with Minimal Labeling

Abstract

In this paper we describe a new approach to detecting and locating brand logos in an image using machine learning methods and synthetic training data. Deep learning methods, particularly the use of Convolutional Neural Networks (CNN), have been very popular for extracting visual information, such as image shapes and objects, from images. A CNN has parameters and configuration information that are learned from training images. To obtain good accuracy usually a large amount of labeled (groundtruthed) images are required for training. Collecting the training images and labeling them can be expensive and time consuming. Methods that include data augmentation, image synthesis, and bootstrapping techniques provide useful alternatives to creating training images. In this paper, we present a logo detection method that requires minimum labeled images. First, we use synthetic images to train a CNN to detect logos. Then, this CNN is used to automatically detect and localize logos from images extracted from the web. Finally, these images are used to train a logo classifier. The combination of the logo detector and the classifier allows us to locate and classify multiple logos in a scene. While existing methods rely on manually labeled images, our method is fully trained with images obtained in an automated manner with minimal human supervision.