Bacterial Raman spectrum is characterized by a weak signal, high similarity, and susceptibility to noise. Its classification using traditional machine learning approaches requires complex spectral preprocessing, and the efficiency is low. In this study, to enhance the accuracy and efficiency of bacterial Raman spectral classification, a one-dimensional convolutional neural network model Raman-net based on dense connection is suggested, which could efficiently complete spectral classification without additional spectral preprocessing. The experimental findings demonstrate that the classification accuracy of Raman-net for 30 bacterial low-signal-to-noise ratios Raman spectra in the Bacteria-ID public data set is 84.26%, which is substantially higher than that of traditional machine learning approaches and comparison approaches. Raman-net attained a classification accuracy of 99.16% for surface-enhanced Raman spectroscopy of 2 Klebsiella pneumoniae susceptible and resistant to carbapenems. This demonstrates that Raman-net can attain remarkable classification findings for ordinary Raman spectroscopy and surface-improved Raman spectroscopy of bacteria without spectral preprocessing, and offers a fast and efficient approach for Raman spectroscopy identification of pathogenic bacteria.