Comparing Deep Learning-based Architectures for Logo Recognition

Logo recognition is a subset of image recognition and has attracted attentions of many researchers due to its specific problem. That is, logo recognition has a wide intra-class and inter-class variability. For example, distinguishing a new edition of a company’s logo and the old one falls into a specific problem that is tailored to logo recognition problem. In this paper, we compare three deep learning-based logo recognition architectures, namely Bianco’s architecture, AlexNet, and Xception. Bianco’s architecture is chosen as a sample of an architecture that includes many preprocessing pipelines including a logo region proposal. Therefore, in this paper we want to investigate whether Bianco’s architecture performs significantly better compared to the others if a logo region proposal is removed. We compare it with other typical deep convolutional neural network architectures such as AlexNet and Xception. Experiments are carried out on the FlickrLogo-32plus, FlickrLogos 27, BrandLogo, and LogoDet-3K. In addition, we also add the curated dataset with hundreds of logo by using Selenium WebDriver. We found out that Bianco’s architecture does not significantly perform better compared to AlexNet, and performs worse compared to Xception. There, we conclude that a logo region proposal is an important preprocessing step in logo recognition.