返回Google圖書搜尋

Improving Face Recognition by Video Spatial Morphing

Armando PadilhaJorge SilvaRaquel Sebastiao

出版	INTECH Open Access Publisher, 2007
ISBN	39026130339783902613035
URL	http://books.google.com.hk/books?id=5BjUoAEACAAJ&hl=&source=gbs_api

註釋This work addressed the applicability of Independent Component Analysis to face recognition through image analysis, namely to the facial authentication problem. We defined and tested the main functionality of a potentially automatic vision system. In fact, some aspects of the process have been performed manually, but it is well known and the literature clearly documents successful automated methods for those operations. The main purpose of the system is to validate the clients' access to restricted areas and to avoid the entrance of impostors, using as few restrictions on the environment as possible, and causing virtually no discomfort on the clients, both by not forcing them to multiple image acquisitions in the database construction phase, and by allowing them to walk naturally (with no specific pose for image capture) in the direction of the entrance. The results obtained have demonstrated the usefulness of the image pre-processing steps described and of the view-morphing techniques to generate virtual frontal views from slightly lateral images of the individuals. The identification of clients was 100% accurate, although it can be observed from the tables in the text that the robustness of the method is not high, as there is a significant number of "near" false-positives. On the other hand, the tests for the rejection of impostors have also proved that this most important goal could be successfully achieved for the data available. Nevertheless, because the two experiments were conducted separately, there are limits to the usability of the results, as is implied by the fact that the computation of the thresholds for identification and for rejection resulted different (that is, one and two standard deviations from mean). This indicates that further work must be done to counter this obvious limitation. One must take into consideration, however, that the building of the database was not thoroughly addressed, as timing constraints in the development phase of the work forced this process to be done very fast. The experience in analysing the results shows that test images of much higher quality can be acquired with only a minor investment in using one or two standard video cameras, and in setting up an adequate illumination and environment conditioning system, in particular to avoid specular reflections in the field of view. Also, the view-morphing point-correspondence problem was solved by manually setting the correspondences and, as previously pointed out, using only a small number of feature points located on a relatively narrow vertical strip of the faces. More accurate methods for matching pairs of points in an image and its reflection can be used in an automatic mode, not only for feature points but also for other points apart from the vertical symmetry line of the face, using epipolar geometry. As a final conclusion, one can say that the results achieved so far are quite encouraging and that the ideas for improving the methodology are well founded and promise a good outcome.