When the convolutional neural network（CNN） is used to classify and identify the vehicle

the influence of the proportions change of vehicle targets in the image on the recognition accuracy of CNN is studied systematically. It is found that if the proportion of targets are too large or too small

the recognition accuracy will be reduced. In view of this effect

the recognition effects of convolution kernel size

stacked convolution structure and batch normalization（BN） layer on multi-proportion vehicle targets are studied and analyzed. Six data sets with different target proportions are established

and the classical CNN model Lenet—5 and the four different CNN models are tested by these data sets. Based on the experimental results of Lenet—5

and comparing the experimental results of different models

the effects of different hierarchical structures on the elimination of target proportions are studied. Among them

the model with the best performance can stabilize the fluctuation of recognition accuracy within 1. 0 %

and the highest recognition accuracy is 97. 33 %. When the model is tested by sample mixing with multiple target proportions

it is found that the CNN model has the highest recognition accuracy for samples with a target proportion of 50 %. The experimental results provide a reference value for the research and application of CNN in vehicle target classification and recognition.