Feature learning boosts network performance

Abstract

Features are an important part of machine learning. Features are often the reduced-dimensional representation of input data, feature calculation, extraction, and fusion directly affect the final result of the network. This thesis proposes a total of 3 feature learning methods, and all of them have achieved good results. In the chapter four of this thesis, the method of video feature extraction is mainly introduced. In this thesis, 3D pre-trained CNN model is used to extract features from continuous video frames, and the feature data is input into the classifier with subnetwork node, and good results are obtained. And the experimental results show that the pre-trained model effectively reduces the training time and improves the testing accuracy. 3D CNN can fully extract the temporal and spatial features of objects. For the subsequent classification accuracy has played a good effect. Chapter three and five are mainly focus on EEG-based signal emotion recognition. EEG data is a one-dimensional signal, but because the human cerebral cortex has different feedback areas for different emotions, the data of EEG data on different electrode channels are spatially related. And different channels of data have a positive impact on different emotional states. Therefore, this chapter improves the accuracy of the test results by converting the one-dimensional data of the EEG into a two-dimensional matrix and adding weights to the corresponding electrodes to extract spatial features and fuse them with frequency-domain features. Since the EEG signal is continuous in time, the time-domain features also have a great effect on the final sentiment classification result. Therefore, the fifth chapter adds the extraction of time-domain features based on the fourth chapter. The final accuracy of emotion recognition has been significantly improved.