Logo Recognition by Combining Deep Convolutional Models in a Parallel Structure

Abstract

In this paper, a new approach is proposed for logo recognition using deep convolutional neural networks. Precise recognition of logos is of high importance in several applications like intelligent traffic control systems and copyright infringement. To enhance the efficiency of logo recognition, we have employed several strategies. In the first strategy, pre-trained deep models are employed for deep feature extraction and classification using the Support Vector Machine (SVM) classifiers. In the second strategy, existing pre-trained deep models are modified for logo recognition after transfer learning and fine-tuning. Finally, fine-tuned models are employed in a parallel structure to enhance the efficiency of logo recognition. We tested the proposed structures with Logos32-plus dataset and the results showed that combining fine-tuned deep models using a voting algorithm gives rise to the best recognition rate of 98.4%. The comparison of results of the proposed structure with a state-of-art deep approach for logo recognition shows the efficiency of the proposed approach.