A brain-computer interface for rapid image searching

Abstract

Even though there have been great advances in computer vision systems, no system has come close to replicating the complexity of the human vision system for object detection. Humans can recognize objects of interest at a glance, even when the objects are shown under different lighting and different angles. The recognition of a target object evokes an identifiable brain activity pattern in an individual, which can be recorded using electroencephalography (EEG). This pattern can be used to increase the efficiency of object detection by using the human vision system for object recognition, and computer processing power to analyse the EEG data and determine whether an object of interest was shown. The aim of this project is to implement a brain-computer interface (BCI) to decode EEG data and determine objects of interest from a series of images shown at a high rate by using rapid serial visual presentation (RSVP). An overview of the system would compose of a stimulus consisting of a series of images containing both target and non-target images. A participant would be subjected to a stimulus and the EEG data would be recorded and used to classify the images shown as target or non-target images by using features extracted from the EEG data to train a classifier. In this project, a stimulus was implemented and data synchronised with the stimulus was recorded from eight subjects. The stimulus consisted of images shown at a rate of five images per second using RSVP. The recorded data was then processed and different feature extraction methods were used to classify the data into target or non-target images. The different feature extraction methods analysed are the decimation method, the all points from t-test result (APT) method, consecutive points from t-test result (CPT) method and the mean of consecutive points from t-test result (MCPT) method. A Fisher linear discriminant analysis (LDA) classifier was used and provided positive results, where the best performing feature extraction method proved to be the decimation method. This method provided a target detection rate of 75 per cent and non-target detection rate of 86 per cent.