Improving Traffic Sign Recognition by Using Wavelet Convolutional Neural Network

Abstract

Traffic sign recognition (TSR) considered as a challenging subject in image processing for many years. Nowadays, after achievements in processing power of processors and easily accessible datasets, many researches has been done by using convolutional neural networks (CNN) In many applications including TSR. CNN is a popular deep learning method that has a reasonable functionality in image classification and pattern recognition.  Important factors in performance of a CNN can be written as follows: accuracy, efficiency and the precision. Therefore, in this paper we try to achieve better results in these important factors. In this particular usage, we must consider complexity and processing time of whole procedure to be in an acceptable range. As will be shown in the following sections, because of wavelet characteristics, we will use wavelet in CNN to improve performance of it in two parts of its structure: In first step, we use wavelet convolutional neural network (wCNN) instead of CNN. Therefore, wavelet transform function is replaced in the convolutional layers of CNN. In next step, wavelet convolutional wavelet neural network (wCwNN) is suggested, so that fully connected neural network (FCNN) of wCNN and CNN is changed by wavelet neural network (wNN). So, we can compare results of these methods (wCNN and wCwNN) with CNN. We can obviously see that, accuracy results is about 4 percent better than CNN and the mean square error and the rate of error are decreased. These results achieved by increasing the calculations of the CNN algorithm.