A separable convolutional neural network for vehicle type recognition

The traditional vehicle type recognition algorithm has a low image recognition rate for various vehicle types on diverse road conditions and is prone to being affected by shooting distance, light intensity, and weather. To address these problems, a new separate convolutional neural network structure was proposed to automatically classify the images of different vehicle types based on the deep learning TensorFlow framework and the classical GoogLeNet-based network model. Experimental results on the data sets of BIT-Vehicle and Cars-196 show that, compared with the traditional HOG_BP algorithm and convolutional neural network model, the decomposed convolutional neural network has a higher recognition rate for the same difficult vehicle images, and its average accuracy rate reaches 96.30%. In addition, the adjustment of hyperparameters in the network ensures that the parameters, such as weight and bias amount, are more efficient and reasonable when constantly updated.