Automatic Frequency Band Selection for Illumination Robust Face Recognition

Ekenel, Hazım Kemal; Stiefelhagen, Rainer

doi:10.1109/icpr.2010.658

Cited by 6 publications

(1 citation statement)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Context-based approaches make use of the geometric qualities of the reference points on the face. In other studies, however, methods that divide the face into sub-regions to alleviate disruptive effects like illumination and change of expression have been used [6, 25]. Hence, the advantages of both methods have been combined.…”

Section: Introductionmentioning

confidence: 99%

Embedded Face Detection and Recognition

Günlü

2012

International Journal of Advanced Robotic Systems

View full text Add to dashboard Cite

The need to increase security in open or public spaces has in turn given rise to the requirement to monitor these spaces and analyse those images on‐site and on‐time. At this point, the use of smart cameras ‐ of which the popularity has been increasing ‐ is one step ahead. With sensors and Digital Signal Processors (DSPs), smart cameras generate ad hoc results by analysing the numeric images transmitted from the sensor by means of a variety of image‐processing algorithms. Since the images are not transmitted to a distance processing unit but rather are processed inside the camera, it does not necessitate high‐ bandwidth networks or high processor powered systems; it can instantaneously decide on the required access. Nonetheless, on account of restricted memory, processing power and overall power, image processing algorithms need to be developed and optimized for embedded processors. Among these algorithms, one of the most important is for face detection and recognition. A number of face detection and recognition methods have been proposed recently and many of these methods have been tested on general‐purpose processors. In smart cameras ‐ which are real‐life applications of such methods ‐ the widest use is on DSPs. In the present study, the Viola‐Jones face detection method ‐ which was reported to run faster on PCs ‐ was optimized for DSPs; the face recognition method was combined with the developed sub‐region and mask‐based DCT (Discrete Cosine Transform). As the employed DSP is a fixed‐point processor, the processes were performed with integers insofar as it was possible. To enable face recognition, the image was divided into sub‐ regions and from each sub‐region the robust coefficients against disruptive elements ‐ like face expression, illumination, etc. ‐ were selected as the features. The discrimination of the selected features was enhanced via LDA (Linear Discriminant Analysis) and then employed for recognition. Thanks to its operational convenience, codes that were optimized for a DSP received a functional test after the computer simulation. In these functional tests, the face recognition system attained a 97.4% success rate on the most popular face database: the FRGC

show abstract

Section: Introductionmentioning

confidence: 99%