A Deep Learning Models for Blind Guidance by Integrating CNN and ELM

Sign recognition is a challenging task but has many important applications. Blind guidance system in one of them. Besides accuracy, which is one of the main metrics as in many other applications, blind guidance system demands real-time response in addition to make the application practical in real word. In this paper, we propose to combine ELM classifier with Deep Convolutional Features, which are obtained from some popular CNNs including AlexNet, GoogLeNet and ResNet, etc., to reduce both training time and prediction time and meanwhile maintain the same level of accuracy comparing to a transfer learning scheme with a single hidden dense layer. Experiments on toilet sign images demonstrate that the proposed method may perform much faster and maintain comparable accuracy. Performance comparisons are also investigated with different CNNs and hyper-parameters.