Image analysis – Pattern recognition – Classification
Reexamination Certificate
1999-03-31
2002-07-16
Mancuso, Joseph (Department: 2623)
Image analysis
Pattern recognition
Classification
C382S279000
Reexamination Certificate
active
06421463
ABSTRACT:
FIELD OF THE INVENTION
This invention relates generally to image processing systems and more particularly to systems for detecting objects in images.
BACKGROUND OF THE INVENTION
As is known in the art, an analog or continuous parameter image such as a still photograph or a frame in a video sequence may be represented as a matrix of digital values and stored in a storage device of a computer or other digital processing device. When an image is represented in this way, it is generally referred to as a digital image. It is desirable to digitize an image such that the image may be digitally processed by a processing device.
Images which illustrate items or scenes recognizable by a human typically contain at least one object such as a persons face, an entire person, a car, etc . . . Some images, referred to as “cluttered” images, contain more than one object of the same type and/or more than one type of object. In a single image or picture of a city street, for example, a number of objects such as people walking on a sidewalk, street signs, light posts, buildings and cars may all be visible within the image. Thus, an image may contain more than one type or class of object (e.g. pedestrians as one class and cars as a different class) as well as multiple instances of objects of the same type (e.g. multiple pedestrians walking on a sidewalk).
As is also known, object detection refers to the process of detecting a particular object or a particular type of object contained within an image. In the object detection process, an object class description is important since the object detection process requires a system to differentiate between a particular object class and all other possible types of objects in the rest of the world. This is in contrast to pattern classification, in which it is only necessary to decide between a relatively small number of classes.
Furthermore, in defining or modeling complicated classes of objects (e.g., faces, pedestrians, etc . . . ) the intra-class variability itself is significant and difficult to model. Since it is not known how many instances of the class are presented in any particular image or scene, if any, the detection problem cannot easily be solved using methods such as maximum-a-posteriori probability (MAP) or maximum likelihood (ML) methods. Consequently, the classification of each pattern in the image must be performed independently. This makes the decision process susceptible to missed instances of the class and to false positives. Thus, in an object detection process, it is desirable for the class description to have large discriminative power thereby enabling the processing system to recognize particular object types in a variety of different images including cluttered and uncluttered images.
One problem, therefore, with the object detection process arises due to difficulties in specifying appropriate characteristics to include in an object class. Characteristics used to specify an object class are referred to as a class description.
To help overcome the difficulties and limitations of object detection due to class descriptions, one approach to detect objects utilizes motion and explicit segmentation of the image. Such approaches have been used, for example, to detect people within an image. One problem with this approach, however, is that it is possible that an object which is of the type intended to be detected is not moving. Thus, in this case, the utilization of motion would not aid in the detection of an object.
Another approach to detecting objects in an image is to utilize trainable object detection. Such an approach has been utilized to detect faces in cluttered scenes. The face detection system utilizes models of face and non-face patterns in a high dimensional space and derives a statistical model for the a particular class such as the class of frontal human faces. Frontal human faces, despite their variability, share similar patterns (shape and the spatial layout of facial features) and their color space is relatively constrained.
Such an approach, without a flexible scheme to characterize the object class, will not be well suited to provide optimum performance unless the objects such as faces have similar patterns (shape and the spatial layout of facial features) and relatively constrained color spaces. Thus, such an approach is not well-suited to detection of those types of objects, such as pedestrians, which typically have dissimilar patterns and relatively unconstrained color spaces.
The detection of objects, such as pedestrians for example, having significant variability in the patterns and colors within the boundaries of the object can be further complicated by the absence of constraints on the image background. Given these problems, direct analysis of pixel characteristics (e.g., intensity, color and texture) is not adequate to reliably and repeatedly detect objects.
One technique, sometimes referred to as the ratio template technique, detects faces in cluttered scenes by utilizing a relatively small set of relationships between face regions. The set of relationships are collectively referred to as a ratio template and provide a constraint for face detection. The ratio template encodes the ordinal structure of the brightness distribution on an object such as a face. The ratio template consists of a set of inequality relationships between the average intensities of a few different object-regions. For example, as applied to faces, the ratio template consists of a set of inequality relationships between the average intensities of a few different face-regions.
This technique utilizes the concept that while the absolute intensity values of different regions may change dramatically under varying illumination conditions, their mutual ordinal relationships (binarized ratios) remain largely unaffected. Thus, for instance, the forehead is typically brighter than the eye-socket regions for all but the most contrived lighting setups.
The ratio template technique overcomes some but not all of the problems associated with detecting objects having significant variability in the patterns and colors within the boundaries of the object and with detection of such objects in the absence of constraints on the image background.
Nevertheless, it would be desirable to provide a technique to reliably and repeatedly detect objects, such as pedestrians, which have significant variability in patterns and colors within the boundaries of the object and which can detect objects even in the absence of constraints on the image background. It would also be desirable to provide a formalization of a template structure in terms of simple primitives, a rigorous learning scheme capable of working with real images, and also to provide a technique to apply the ratio template concept to relatively complex object classes such as pedestrians. It would further be desirable to provide a technique and architecture for object detection which is trainable and which may also be used to detect people in static or video images of cluttered scenes. It would further be desirable to provide a system which can detect highly non-rigid objects with a high degree of variability in size, shape, color, and texture and which does not rely on any a priori (hand-crafted) models or on changes in position of objects between frames in a video sequence.
SUMMARY OF THE INVENTION
In accordance with the present invention, an object detection system includes (a) an image preprocessor for moving a window across the image and a classifier coupled to the preprocessor for classifying the portion of the image within the window. The classifier includes a wavelet template generator which generates a wavelet template that defines the shape of an object with a subset of the wavelet coefficients of the image. The wavelet template generator generates a wavelet template which includes a set of regular regions of different scales that correspond to the support of a subset of significant wavelet functions. The relationships between different regions are expressed as constraints on the values of the wavele
Oren Michael
Papageorgiou Constatine P.
Poggio Tomaso
Sinha Pawan
Daly, Crowley & Mofford LLP
Dastouri Mehrdad
Mancuso Joseph
Massachusetts Institute of Technology
LandOfFree
Trainable system to search for objects in images does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Trainable system to search for objects in images, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Trainable system to search for objects in images will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2913369