Traffic sign recognition based on multi-scale feature fusion and extreme learning machine

Feature extraction is performed on a convolutional neural network using only the same size image and convolution kernel, which results in incomplete extraction features. In the traffic sign recognition, the continuous change of the position between the vehicle camera and the traffic sign affects the recognition accuracy of the traffic sign. A traffic sign recognition method based on multi-scale feature fusion and extreme learning machine is proposed. Firstly, the network model adapted to three different size images is pre-trained as the initial model of the experiment. Then, the three network models are combined to construct a multi-scale convolutional neural network, and the parameters of the three pre-trained networks are cascaded to the full connection of the fusion model to train the fully connected layer of the fusion model, and the network parameters are updated by the random gradient descent algorithm. Finally, the fusion model is used as the feature extractor to extract features, which are sent to the extreme learning machine(ELM) to realize traffic sign recognition. The experimental results show that the recognition accuracy of the multi-scale feature fusion and extreme learning machine combined network is 99.23% and the recognition speed is 46 ms. Compared with the pre-trained network, the classification accuracy of the network is increased by 2.35%, 3.22% and 3.74%, respectively. Therefore, the multi-scale feature fusion method can effectively extract the feature information of traffic sign images, ELM classifier can improve classification accuracy and classification time, and meet the requirements of traffic sign recognition accuracy and real-time.