A Saliency Guided Shallow Convolutional Neural Network for Traffic Signs Retrieval

Abstract

As one of the important parts of road infrastructure, traffic signs provide vital information for road users. Achieving efficient traffic signs retrieval greatly contributes to the intelligent analysis on big traffic data. In this paper, we propose a saliency guided shallow convolutional neural network (CNN) for traffic signs accurate and fast retrieval. Firstly, by unifying deep saliency and hashing learning in a single architecture, the proposed CNN model performs joint learning in a point-wise manner, which is scalable on large-scale datasets. Then, deep saliency features and hashing-like outputs are extracted from traffic sign images with the saliency guided shallow CNN. The binarized hashing-like outputs together with saliency features are used to construct features database. Finally, a coarse to fine similarity measurement is performed by Euclidean distance and Hamming distance to return retrieval results. Experimental results demonstrate the retrieval accuracy of our method outperforms five state-of-the-art methods on GTSRB dataset.