Effective traffic signs recognition via kernel PCA network

The classification of traffic sign images is easily affected by the change of weather, camera angles, occlusion, etc. The traditional image recognition methods not only require high image quality, but also need to find effective features manually. However, the convolutional neural networks can automatically extract high-level, abstract features which are robust to the variations. This paper presents a novel and effective traffic signs recognition approach via the kernel PCA network based on convolutional neural networks. The kernel PCA network uses two-layer convolutional network to extract abstract features, and convolution kernels in each layer are directly calculated by the kernel principal component analysis. After nonlinear mapping and pooling, support vector machines are applied to the final classification. The approach can achieve a high recognition rate on the German traffic signs recognition benchmark dataset containing reliable ground-truth data.