From Signal to Image Then to Feature: Decoding Pigeon Behavior Outcomes During Goal-Directed Decision-Making Task Using Time-Frequency Textural Features

Abstract

Neural information decoding has become a hot issue in the field of brain science and the brain-computer interface. To effectively decode pigeon behavior outcomes during goal-directed decision-making task, a goal-directed experiment based on plus maze is designed and six pigeons are trained to acquire neural signals. Then, continuous wavelet transform (CWT) is employed to obtain the time-frequency images of local field potential (LFP) signals, and a series of textural features based on different texture analysis descriptors are extracted. Finally, textural features are used to decode different behavior outcomes based on random forest (RF). Decoding performances of single-channel LFP signals at different frequency bands, signals containing different numbers of channels, different types of features and classifier parameters are compared and discussed. The results show that the time-frequency textural features can be used to decode the animal behavior outcomes effectively. Specifically, the textural features of the LFP signal in the 40–60 Hz frequency band perform best in pigeon behavior decoding, and its performance is less affected by the number of channels. Different types of texture features have different decoding performances, and it seems that some local jet based gray-level statistic (LJbG) features are more suitable in this study. The performance comparison of different RF parameters shows that better decoding accuracy can be guaranteed when the number of trees is set at 6–8. These results will help us to understand the brain neuronal information processing mechanism of pigeon further.