Logo recognition using CNN features

Abstract

In this paper we propose a method for logo recognition based on Convolutional Neural Networks, instead of the commonly used keypoint-based approaches. The method involves the selection of candidate subwindows using an unsupervised segmentation algorithm, and the SVM-based classification of such candidate regions using features computed by a CNN. For training the neural network we augment the training set with artificial transformations, while for classification we exploit a query expansion strategy to increase the recall rate. Experiments were performed on a publicly-available dataset that was also corrupted in order to investigate the robustness of the proposed method with respect to blur, noise and lossy compression.