Scientists at the Berkeley SETI Research Center are Searching for Extraterrestrial Intelligence (SETI) by a new signal detection method that converts radio signals into spectrograms through Fourier transforms and classifies signals represented by two-dimensional time-frequency spectrums, which successfully converts a signal classification problem into an image classification task. In view of the negative impact of background noises on the accuracy of spectrograms classification, a new method is introduced in this paper. After Gaussian convolution smoothing the signals, edge detection functions are applied to detect the edge of the signals and enhance the outline of the signals, then the processed spectrograms are used to train the deep neural network to compare the classification accuracy of various image classification networks. The results show that the proposed method can effectively improve the classification accuracy of SETI spectrums.