The classification and detection of quasars through analysis of spectrum is a major research topic in astronomy. Quasars are not easy to classify due to its noises and dramatic changes in spectra. On the other side, traditional methods of template matching are not effective in quasar recognition. Therefore, by improving the Convolutional Neural Network (CNN) which is widely used in image classification, this paper makes full use of CNN for its excellent performance in feature extraction and applies it to automatically identify spectra of quasars which are in the form of row vectors. The improved method is used for quasar identification in Sloan Digital Sky Survey (SDSS) published data. All data are averaged filtered and normalized before they are trained and tested. It is found that this method presents excellent results and is very promising for project such as Five-hundred-meter Aperture Spherical radio Telescope (FAST).