Font Size: a A A

A Closed-loop System For Face Retrieval Based On EEG And Computer Vision

Posted on:2017-03-14Degree:MasterType:Thesis
Country:ChinaCandidate:Y WangFull Text:PDF
GTID:2308330485957129Subject:Biomedical engineering
Abstract/Summary:PDF Full Text Request
Computer vision has already achieved great success in face retrieval from large face database. However, its performance suffers from a severe degradation when human face’s expression, pose or illumination varies beyond a certain degree. Human brain has an excellent ability of face recognition which is shown to be very robust to condition changes such as human face’s expression, pose or illumination. Some researchers found that EEG (Electroencephalography) signal contains many ERP (Event Related Potential) components related with human face recognition. It is feasible to implement rapid face retrieval by single trial facial ERP detection. In this thesis, we propose a closed-loop system coupling human brain and computer vision for rapid face retrieval. This system integrates human’s excellent ability of face recognition and computer vision’s fast computing ability efficiently by the brain computer interface based on single trial ERP detection.This closed-loop system consists of three modules, face database, EEG module and computer vision module. During each iteration, face images are presented to subject by rapid serial visual presentation (RSVP) paradigm. EEG module evaluates subject’s most interested candidate images by single-trial ERP detection. Computer vision module retrieves the candidates’most similar images from the face database for the next RSVP. After a few iterations, the face retrieval result is the ranking of all the images in the face database in the last iteration. Retrieval result is estimated by the average precision (AP). Compared with previous closed-loop systems, this is the very first time that closed-loop system is applied to retrieve face images which is relatively difficult and we design a refinement policy implemented by supervised learning (Support Vector Machine, SVM) and unsupervised learning (K-means and Self-organizing Map, SOM).The offline analysis results showed that our supervised target refinement strategy boosted the ratio of target face images in the EEG candidate face images from 69.76% to 96.95% and retrieval performance from 0.8067 to 0.9044. Furthermore, target refinement strategy based on unsupervised learning boosted the ratio of target face images to 95.03% and retrieval performance to 0.9162. Finally, the online human face retrieval experiment verified the closed-loop system’s superior performance.In conclusion, combining the brain computer interface based on EEG with computer vision technology can overcome computer vision’s defects of failing in face recognition when the face has variation in illumination, pose and expression, and achieves relatively fast and accurate face retrieval.
Keywords/Search Tags:Closed-loop system, Brain-Computer Interface, Single trial event related potential(ERP)detection, Face retrieval, Computer vision
PDF Full Text Request
Related items