| Summary: | 碩士 === 國立臺灣大學 === 電子工程學研究所 === 99 === Multichannel neural recording is one of the most important topics in the field of biomedical engineering. This is because there is a need to considerably reduce large amounts of data without degrading the data quality for easy transfer through wireless transmission. Both spontaneous data and evoked data need compression. Spontaneous data has good SNR, and it can be easily detected by a simple threshold. But some further analysis need very high accuracy. Nowadays spike sorting technique cannot afford such high accuracy(95% up). So we still need compression method to get the original data for manual sorting. Evoked data has low SNR. It cannot be easily detected. Original data is the only way to analyze.
Video compression technology is of considerable importance in the field of signal processing. There are many similarities between multichannel neural signals and video signals. In this study, we propose a signal compression method that employs motion vectors (MVs) to reduce the redundancy between successive video frames and between successive channels. We also try different transforms to get the best results. We try DCT (Discrete Cosine Transform), DST (Discrete Sine Transform), 5-3DWT(Discrete Wavelet Transform), Hadamard transform. Finally we discover the best performance is DCT.
Further we utilize the hybrid coding method to get better performance. Although there is no evidence showing that there is correlation or prediction between time domain or cross-experiments. But if we look inside the DCT (Discrete Cosine Transform), there is very highly concentrated in time domain. The energy is compact. Also, cross-experiments figure shows that there are many similarities between different experiments. During this stage, we also change the scan mode from "zig-zag" to straight. This is because there is different energy distribution between nature signal and neural signal. This changing of scan mode contributes 2 db gain without algorithm changing.
Thus, we perform hybrid-coding to reduce the different part of redundancy. In intra-mode, we perform the channel and time domain correlation to reduce the data. In inter-mode, we use cross-experiment redundancy to do the compression. We also try 3 different frame setting to compare. Channel-mode, time-mode, and event-mode has different settings and results. Event-mode utilizes the most redundancy, so it has best performance.
Under CR(compression ratio)=16, the event-mode setting can get SNR=27.8db, compared the previous work there is 4 db improvement.
|
|---|
