Hierarchical CNN for traffic sign recognition

Abstract

The Convolutional Neural Network (CNN) is a breakthrough technique in object classification and pattern recognition. It has enabled computers to achieve performance superior to humans in specialized image recognition tasks. Prior art CNNs learn object features by stacking multiple convolutional/non-linear layers in sequence on top of a classifier. In this work, we propose a Hierarchical CNN (HCNN) which is inspired by a coarse-to-fine human learning methodology. For a given dataset, we introduce a CNN-oriented clustering algorithm to separate classes into K subsets, which are referred to as families. Then, the HCNN algorithm trains K+1 classification CNNs: one CNN for family classification and K dedicated CNNs corresponding to each family for member classification. We evaluate this HCNN approach on the German Traffic Sign Recognition Benchmark (GTSRB), and achieve 99.67% correct detection rate (CDR), which is superior to the best reported results (99.46%) achieved by a single network.