Abstract: In the report, some vehicle-logo samples were collected from the video which is not clear enough, and the amount of the sample was expanded by adding noise to the samples. A three layer convolution neural network was designed to find the suitable parameter and extract the characteristics from the training samples. The information from the training samples was used for the model training and classification. The results indicated that under the condition of illumination change and noise pollution, the accuracy of the method was still high, and the accuracy rate of classification was as high as 99.06% on its own dataset.