The paper presents a Multi-Column Deep Neural Network (MCDNN) that won the German Traffic Sign Recognition Benchmark, achieving a recognition rate of 99.46%, surpassing human performance. The method uses a fast, GPU-based implementation of a Deep Neural Network (DNN) that does not require pre-wired feature extractors, which are learned in a supervised manner. By combining multiple DNNs trained on differently preprocessed data, the MCDNN enhances recognition performance and robustness to variations in contrast and illumination. The DNN architecture consists of convolutional and max-pooling layers, with the final output being a softmax classification. The MCDNN is formed by averaging the outputs of multiple DNN columns, each trained on different preprocessing methods. The paper details the training process, data preprocessing techniques, and experimental results, demonstrating the effectiveness of the MCDNN in traffic sign classification.The paper presents a Multi-Column Deep Neural Network (MCDNN) that won the German Traffic Sign Recognition Benchmark, achieving a recognition rate of 99.46%, surpassing human performance. The method uses a fast, GPU-based implementation of a Deep Neural Network (DNN) that does not require pre-wired feature extractors, which are learned in a supervised manner. By combining multiple DNNs trained on differently preprocessed data, the MCDNN enhances recognition performance and robustness to variations in contrast and illumination. The DNN architecture consists of convolutional and max-pooling layers, with the final output being a softmax classification. The MCDNN is formed by averaging the outputs of multiple DNN columns, each trained on different preprocessing methods. The paper details the training process, data preprocessing techniques, and experimental results, demonstrating the effectiveness of the MCDNN in traffic sign classification.