Sorting Overlapping Spikes Based on Log-Mel Spectrogram and Convolutional Neural Networks

Sorting overlapping spikes is a key problem in spike classification. The overlapping spikes not only affect the accuracy of spike classification, but also interfere with correct information decoding. The template matching method is the main method to distinguish overlapping spikes, and its accuracy is also better than other methods. However, with the increase of collected data, the traditional template matching method consumes a lot of time and manpower to distinguish overlapping spikes. Therefore, this paper proposes a new method based on template to automatically distinguish overlapping spikes using Log-Mel spectrogram and convolution neural network (Mel_CNN). Firstly, make overlapping spike templates according to the existing spike categories. Secondly, use the spike and overlapping spike templates to train Mel_CNN. Finally, test the trained network with real spike and overlapping spike, The trained network can not only classify spikes into correct categories, but also identify overlapping spikes. The Mel_CNN method proposed in this paper not only outperforms the traditional template matching method in time, but also shows that the average accuracy of Mel_CNN method can reach 95.74%, so the method can effectively distinguish overlapping spikes. And validating the model with different metrics, the results are stable on datasets with different noise levels, suggesting out model has a strong robustness.