Abstarct: In recent years, the real-time classification of vehicle types has been one of the most interesting and widely used topics due to its application in the control and analysis of road traffic. Deep neural networks are one of the most attractive methods for classifying vehicles due to their enormous capabilities in classification issues. One of the problems of these structures is the high computational load of the network, which slows down the process and makes it difficult to implement for real-time applications. In this paper, a new convolutional-baxsed network for vehicle type recognition is introduced, which performs very well in both CPU and GPU platforms and operates in real-time. The innovations of this structure include both the network architecture and the structural elements that make it up, as well as the method of learning and implementing the back-propagation learning algorithm. In common classifiers, feature extraction takes place only in the convolution laxyer, while in the proposed structure, feature extraction takes place at three locations in the network. The proposed structure consists of five main blocks, each with a convolution laxyer, a drop out laxyer, an SE block and a proposed wavelet-baxsed pooling laxyer, and a batch normalizer. Another innovation is the proposal of a method for updating the weights in the back-propagation learning algorithm. The use of this method increased the accuracy of the proposed structure and similar structures in various optimizers. The proposed method is evaluated on the IRVD dataset, which was collected by the authors, and the MIO-TCD dataset. The proposed structure on IRVD achieves 42 ms recognition time with about 99.59% accuracy on CPU, which is optimal both in terms of accuracy and recognition time, compared to the tested models (which are common models). The proposed structure on the MIO-TCD data set is also 96.59% accurate. As will be shown in the Experimental Results section, this structure has achieved the best results in terms of classification criteria compared to conventional CNNs.