Image analysis – Applications – Target tracking or detecting
Reexamination Certificate
1999-01-11
2002-10-08
Boudreau, Leo (Department: 2721)
Image analysis
Applications
Target tracking or detecting
C382S118000, C348S169000
Reexamination Certificate
active
06463163
ABSTRACT:
TECHNICAL FIELD
The invention relates generally to image pattern detection and recognition and more particularly to a system and a method for face detection.
DESCRIPTION OF THE RELATED ART
Vast amounts of digital information, including images, are currently available via the Internet or other electronic databases. Unlike text information, content-based searches of images containing a picture of a target object from these databases is a challenging task. The difficulty arises from the fact that the pictures of the target objects in the stored images, in general, are not identical. Moreover, the target object depicted in one image may have a different orientation and size from the identical target object depicted in another image.
Face detection technology is being advanced, in part, to assist in the development of an image retrieval system that can overcome the above-described difficulty. Face detection is a process of determining whether a picture of a human face is present in an input image and, if so, accurately determining the position(s) of the face(s) within the input image. A face detector is designed to scan the input image to detect human faces that may be depicted in the image, regardless of the size of the human faces. There are two prominent approaches to face detection, a “neural network-based” approach and an “eigenface” approach.
The neural network-based approach utilizes, as the name suggests, a neural network to detect a human face in an input image. The fundamental ideal of the neural network-based approach is to design a neural network that accepts an N×M image block and outputs a binary answer indicating a positive or a negative detection of a human face within the image block. The neural network is trained using a large database of training image blocks. The training image blocks are a mixture of face images and non-face images. The training image blocks are typically preprocessed before being input to the neural network. The preprocessing may include removing the DC component in the image block and normalizing the image block. After the neural network has been trained, an input image can be analyzed by the neural network during an on-line detection procedure in which N×M image blocks of the input image are preprocessed in the same manner as the training image blocks.
A publication by Henry A. Rowley, Shumeet Baluja and Takeo Kanade, entitled “Neural Network-Based Face Detection,”
IEEE Transactions on Pattern Analysis and Machine Intelligence
, Vol. 20, No. 1, pages 23-28, January 1998, describes a neural network-based face detection system. The face detection system of Rowley et al. utilizes a modified version of the standard neural network-based face detection approach. In particular, the Rowley et al. face detection system uses the training image blocks that have been erroneously detected as containing a human face to further train the neural network in a “bootstrap” manner. The publication asserts that the bootstrapping technique reduces the number of training image blocks that are required to sufficiently train the neural network. In addition, the Rowley et al. face detection system includes neutralization of illumination differences in image blocks during the preprocessing procedure by removing the best-fit linear function from the image blocks.
The eigenface approach involves calculating the best linear basis, or principal eigenvector components, called “eigenfaces” to approximate a set of training faces. These basis vectors are then used as convolution kernels for matched filtering to detect human faces in an input image. U.S. Pat. No. 5,710,833 to Moghaddam et al. describes an apparatus for detection and recognition of specific features in an image using an eigenface approach. The Moghaddam et al. apparatus utilizes all eigenvectors, not just the principal eigenvector components. The use of all eigenvectors is intended to increase the accuracy of the apparatus to detect complex multi-featured objects.
Although the conventional face detection systems operate well for their intended purposes, what is needed is a face detection system and a method of detecting faces that increase face detection performance in terms of speed and accuracy.
SUMMARY OF THE INVENTION
A face detection system and a method of pre-filtering an input image for face detection utilize a candidate selector that selects candidate regions of the input image that potentially contain a picture of a human face. The candidate selector operates in series with an associated face detector that verifies whether the candidate regions do contain a human face. The pre-filtering operation performed by the candidate selector screens out much of the input image as regions that do not contain a human face. Since only the candidate regions are then processed by the face detector, the operation of the candidate selector reduces the amount of computational processing that must be performed by the face detector.
In the preferred embodiment, the candidate selector includes a linear matched filter and a non-linear filter that operate in series to select the candidate regions of the input image. The linear matched filter operates to select image regions that have highly similar image patterns when compared to a face template. The linear matched filter includes a linear correlator and a processing module. The linear correlator performs a linear correlation on the input image using a filtering kernel to derive a correlation image. The filtering kernel is a numerical representation of the face template. The filtering kernel is calculated during a training period, or a non-face detecting period, by a filtering-kernel generator. Preferably, the linear correlation is performed in the discrete cosine transform (DCT) domain, but other approaches are available. The correlation image is then examined by the processing module. The processing module is configured to select temporary candidate regions of the input image using a decision rule. The decision rule dictates that only image regions that are positioned about a local maximum in the correlation image and have pixel correlation values that are greater than a threshold correlation value are to be selected. The temporary candidate regions are then transmitted to the non-linear filter.
The non-linear filter operates to determine whether the temporary candidate regions should be deemed as the candidate regions. The non-linear filter examines contrast values within certain regions of a temporary candidate region, seeking a contrast pattern that is characteristic of eyes of a human face. High contrast values at these regions equate to a likelihood that an image region contains a human face. In one embodiment, the non-linear filter includes three contrast calculators and a decision module. The contrast calculators compute contrast values for particular upper segments of an image region. The first contrast calculator computes a contrast value for an upper-half segment of the image region. The second contrast calculator computes contrast values for two upper quadrants of the image region. Thus, the first and second contrast calculators are dedicated to the top fifty percent of a temporary candidate region. The third contrast calculator computes contrast values for three adjacent segments that define a portion of the upper-half segment, e.g., the top thirty-three percent of the temporary candidate region. These contrast values are transmitted to the decision module. The contrast values computed by the second and third contrast calculators are compared to a threshold contrast value. If these values exceed the threshold contrast value, the image region is deemed to be a candidate region and is transmitted to the face detector. In an alternative configuration, the three contrast calculators may be embodied in a single contrast calculator.
The face detector of the system may utilize a neural network-based approach, an eigenface approach or any other known technique to detect a human face in the candidate regions. In the preferred embodiment, the face detector
Boudreau Leo
Choobin M B
LandOfFree
System and method for face detection using candidate image... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with System and method for face detection using candidate image..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and System and method for face detection using candidate image... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2935453