Functional and effective connectivity based features of EEG signals for object recognition

Abstract

Classifying different object categories is one of the most important aims of brain-computer interface researches. Recently, interactions between brain regions were studied using different methods, such as functional and effective connectivity techniques. Functional and effective connectivity techniques are applied to estimate human brain areas connectivity. The main purpose of this study is to compare classification accuracy of the most advanced functional and effective methods in order to classify 12 basic object categories using Electroencephalography (EEG) signals. In this paper, 19 channels EEG signals were collected from 10 healthy subjects; when they were visiting color images and instructed to select the target images among others. Correlation, magnitude square coherence, wavelet coherence (WC), phase synchronization and mutual information were applied to estimate functional cortical connectivity. On the other hand, directed transfer function, partial directed coherence, generalized partial directed coherence (GPDC) were used to obtain effective cortical connectivity. After feature extraction, the scalar feature selection methods including T-test and one-sided-anova were applied to rank and select the most informative features. The selected features were classified by a one-against-one support vector machine classifier. The results indicated that the use of different techniques led to different classifying accuracy and brain lobes analysis. WC and GPDC are the most accurate methods with performances of 80.15% and 64.43%, respectively.

Keywords: Brain connectivity; Effective connectivity; Electroencephalography (EEG); Functional connectivity; Object recognition.

Fig. 1

All 12 categories. All Categories including animal, flower, fruit, transportation device, body organ, clothing, food, stationery, building, electronic device, doll and jewel are represented

Fig. 2

The order of images in the experiment. Target cue is presented on the screen for 1 s, then after 900 ms. showing a black screen, each image is displayed for 700 ms with an 800 ms delay between consecutive images

Fig. 3

The order of EEG Recording channels. EEG signals were recorded from 19 channel electrodes

Fig. 4

Optimization of SVM parameters. Searching for the best values for regularization constant and kernel argument is done by finding the best classification accuracy for WC method. The best choice for regularization constant and kernel argument is (6.30, 6.30)

Fig. 5

The optimal number of features based on t-test criterion. The optimal number of features is shown as 300 because of the best test and train data classification accuracies on (300, 79.44) and (300, 85), respectively

Fig. 6

The averaged confusion matrix using WC and one-sided-anova. The best classification accuracy is related to category of animals with accuracy percentage of 100

Fig. 7

The averaged confusion matrix using WC and t-test. The best classification accuracy is allocated to category of food

Fig. 8

The averaged brain lobes connectivity matrix for all subjects using WC and t-test. The most connectivity was observed in temporal–frontal and central–frontal lobes

Fig. 9

The averaged brain lobes connectivity matrix using all functional feature extraction techniques and feature selection based on t-test criterion. Connectivity had more strength between frontal–frontal, frontal–central and frontal–temporal lobes

Fig. 10

The averaged brain lobes connectivity matrix using all effective feature extraction techniques and feature selection based on t-test criterion. Connectivity had more strength between parietal–parietal, occipital–occipital and temporal–temporal lobes