Few-shot traffic sign recognition with clustering inductive bias and random neural network. (April 2020)

• Record Type: Journal Article
• Authors: Zhou, Shichao; Deng, Chenwei; Piao, Zhengquan; Zhao, Baojun
• Is Part Of: Pattern recognition. Volume 100(2020:Apr.)
• Publication Date: 2020-04
• Subjects: Traffic sign recognition -- Few-shot learning -- Clustering -- Randomization,; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2019.107160
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A novel generative feature learning framework for TSR is proposed with clustering inductive bias. Clustering-oriented feature mapping is learned based on a novel random neural network. Computationally efficient feature mapping can be achieved with a fast Gaussian random projection. Abstract: Reliable and fast traffic sign recognition (TSR) that locates the traffic sign from an image and then estimates its category is a crucial perception function for Advanced Driver Assistance Systems (ADAS) of autonomous vehicles. Most of the popular deep convolutional neural networks (DCNNs) based TSR techniques advocate discriminative feature learning for traffic signs against their appearance variability. However, such feature learning scheme may suffer from the diversity of traffic signs categories, especially when samples within each category are limited for model training (i.e., few-shot learning). Here, we present a generative feature learning based TSR network with well generalization capacity and high computational efficiency. Instead of relying on large amounts of supervision to learn discriminative features, our method devotes to learn common but unique properties of class-specific traffic signs with few training samples. Specifically, we combine clustering inductive bias with a random neural network, and then exploit computational advantages offered by a fast random projection algorithm. Experiments on two TSR benchmarks illustrate that our method achieves comparable or higher recognition accuracy than state-of-the-art DCNN-based methods with less training data and inference time consumption.