This paper analyses different types of artifacts that appear in neural recording experiments and thus a method is proposed to detect and remove artifacts as a part of preprocessing procedures before information decoding. Through modeling and data analysis, we reason that artifacts have different spectrum statistics compared with field potentials and spikes and the frequency bands of 150-400 Hz and >5 kHz are the most prospective regions to detect artifacts. A synthesized database based on recorded neural data and manually labeled artifacts has been built to allow quantitative evaluations of the proposed algorithm. Testing results have shown that over >80% positive detection ratio is achievable for artifacts with magnitude comparable to neural spikes. Quantitative signal-to-distortion ratio (SDR) simulation has shown that it is possible to have 10-30dB SDR improvement at waveform segments that contain artifacts.