DESCRIPTION
Method and System for Detecting Specific Object in Image
Technical Field
The present invention relates generally to the field of computer vision and pattern recognition, and more particularly to detecting arbitrarily oriented objects in images.
Background Art
All of the applications where computer vision is used, face detectionpresents an extremely difficult challenge . For example, in images acquired by surveillance cameras, the lighting of a scene is usually poor and uncontrollable, and the cameras are of low quality and usually distant from potentially important parts of the scene . Significant events are unpredictable . Often, a significant event is people entering a scene. People are typically identifiedby their faces . The orientation of the faces in the scene is usually not controlled. In other words, the images to be analyzed are substantially unconstrained.
Face detection has a long and rich history. Some techniques use neural network systems, see Rowley et al ., "Neural network-based face detection," IEEE Patt. Anal. Mach. Intell., Vol. 20, pp. 22-38, 1998. Others use Bayesian statistical models, see Schneiderman et al., "A statistical method for 3D object
detection applied to faces and cars, " Computer Vision andPattern Recognition, 2000. While neural network systems are fast and work well, Bayesian systems have better detection rates at the expense of longer processing time.
The uncontrolled orientation of faces in images poses a particularly difficult detection problem. In addition to Rowley et al . and Schneiderman et al . , there are a number of techniques that can successfully detect frontal upright faces in a wide variety of images . Sung et al . , in λΕxample-based learning for view based face detection," IEEE Patt. Anal. Mach. Intell . , volume20, pages 39-51, 1998, described an example-based learning technique for locating upright, frontal views of human faces in complex scenes . The technique models the distribution of human face patterns by means of a few view-based "face" and "non-face" prototype clusters . At each image location, a different feature vector is computed between the local image pattern and the distribution-based model. A trained classifier determines, based on the difference feature vector, whether or not a human face exists at the current image location.
While the definition of "frontal" and "upright" may vary from system to system, the reality is that many images contain rotated, tilted or profile faces that are difficult to detect reliably.
Non-upright face detection was described in a paper by Rowley
et al., "Rotation invariant neural network-based face detection, " Proceedings ofthe IEEEConference on ComputerVision and Pattern Recognition, pages 38-44, 1998. That neural network based classifier first estimated an angle of rotation of a front-facing face in an image. Only the angle of rotation in the image plane was considered, i.e., the amount of rotation about z-axis . Then, the image was rotated to an upright position, and classified. For further detail, see, Baluja, et al., U.S. Patent No. 6,128,397, "Method for finding all frontal faces in arbitrarily complex visual scenes," October 3, 2000.
Figure 1 show the steps of the prior art face detector . A rotation of a front facing face in an image 101 is estimated 110. The rotation 111 is used to rotate 120 the image 101 to an upright position. The rotated image 121 is then classified 130 as either a face or a non-face 131. That method only detects faces with in-plane rotation. That method cannot detect faces having an arbitrary orientation in 3D.
Therefore, there is a need for a system and method that can accurately detect arbitrarily oriented objects in images.
Disclosure of Invention
The invention provides a method for detecting a specific object in an image. An orientation of an arbitrary object in an image is determined and one of a plurality orientation and object
specific classifiers is selected according to the orientation. The arbitrary object is classified as a specific object with the selected orientation and object specific classifier.
Brief Description of Drawings
Figure 1 is a flow diagram of prior art method for detecting in in-plane, rotated, front facing face;
Figure 2 is a block diagram of a system and method for detecting an object having an arbitrary orientation;
Figures 3A-3D are block diagrams of rectangular filters used by the invention; and
Figure 4A-4D are block diagrams of rectangular filters arranged diagonally.
Best Mode for Carrying Out the Invention
System Structure
Figure 2 shows a system 200 for detecting a specific obj ect having an arbitrary orientation in an image 201 according to the invention. By orientation, we mean a rotation about any or all of the three major (x, y, and z axes), for example pitch, yaw, and roll, with respect to an image plane at the instant the image
201 is acquired. We distinguish our orientation from the single rotation about the z-axis of the prior art. In one example application, the objects detected in the images are faces, however, it should be understood that other arbitrarily oriented objects can also be detected. It should also be understood, from the perspective of the camera, that the same method can be used determine an orientation of a camera with respect to a fixed object.
The system 200 includes an orientation classifier 210, a classifier selector 220, and an orientation and object specific classifier 230, connected to each other. The system 200 takes as input an image including an arbitrary object 201 and outputs a detected specific object 231 in the image 201. The classifier selector 220 uses an orientation class 211 and a set of orientation and object specific classifiers 212 to output a selected classifier 221.
In a preferred embodiment, the image is partitioned into detection windows or NNpatches" of various sizes, for example, the entire image, four windows, each H of the image, and so forth.
System Operation
During operation, a method first determines 210 an orientation class 211 of the arbitrary object in the image 201. An orientation
and object specific classifier 221 is selected 220 from a set of orientation and object specific classifiers 212 according to the determined orientation class 211 of the arbitrary object in the image 201. The arbitrary object is then classified 230 as a specific object 231 with the selected orientation and object specific classifier 221.
The classifiers can be any known classifier, e.g., Bayesian, neural network based, support vector machine, decision tree, etc.
Orientation Classifier
The orientation classifier 210 is a multi-class classifier trained on onlypositive image samples of objects to be classified, e.g., faces. Positive image samples means that each image sample is an example of the specific object. The positive samples include the specific object in any or all of the possible orientations on the three major axes. Samples in the possible orientations of the arbitrary object with respect to the image plane at the instant the image is acquired are grouped in classes, e.g., each orientation class includes specific objects having an orientation within a predetermined range of degrees pitch, yaw and roll for the class. The positive samples are labeled according to orientation class. Every arbitrary object input to the orientation classifier is classified as having a
particular orientation class. If the arbitrary object is not the specific obj ect, the output 211 of the orientation classifier 210 is a random orientation class.
In a preferred embodiment, the orientation classifier uses a decision tree as described by Quinlan, "Improved use of continuous attributes in C4.5," Journal of Artificial Intelligence Research 4, 77-90, 1996 incorporated herein by reference .
Each node function is a filter from a set of rectangle filters, described below, and there is no pruning. Every node of the tree is split until amaximumleaf depth is attainedor the leaf contains examples of only one node.
Orientation and Object Specific Classifiers
Each classifier in the set of orientation and object specific classifiers 212 is a binary classifier for detecting the specific obj ect at a particular orientation in the detection window. Each classifier in the set is trained on specific objects in one of the orientation classes. The selected classifier 221 is the orientation and object specific classifier trained on specific objects in the orientation class 211 output by the orientation classifier 210.
Each of the orientation classes described above can include image samples in a range of degrees of rotation about one or all of the three major axes, for example, in a preferred embodiment, the range can be ^f15°. The filters we describe below can be
rotatedby 90°. Therefore, each orientation and object specific
classifier can also be rotated by 90°. As an example, a frontal
face detector trained at 0°, can be rotated about the z-axis
to yield detectors for 90°, 180° and 270° as well. The same
rotations can be performed on classifiers trained at 30° and
60° respectively. Taking into account the range of 4^15° in this example, all frontal-rotation orientation classes canbe covered by 12 classifiers, as opposed to 360 classifiers. Similar classifiers can be trained for other orientations.
Filters, Features, and Classifiers
Formally, operations with our filters, features and classifiers of the preferred embodiment are defined as follows, see U.S. PatentApplication S/N 10/200, 726, "Object Recognition System, " filed by Viola et al . , on July 22, 2002, incorporated herein by reference. An image feature hd{x) is
assigned a weight atj or β
j according to hJx) ,
where a filter fj (x) is a linear function of an image x, i.e., a detection window, and θj is a predetermined filter threshold
value . An cumulative sum C (x) is
assigned a value 1 or 0 according to ,
where h3 are multiple features of the image x, and T is a predetermined classifier threshold.
The selected orientation and object specific classifier 230 rejects the arbitrary object 201 when an accumulated score is less than the classifier threshold and classifies the arbitrary object as the specific object 231 when the cumulative score is greater than the classifier threshold.
In the preferred embodiment, our system uses rectangle filters as described by Viola et al., above. Figures 3A- D show three types of known rectangle filters that the invention can use. The value of a two-rectangle filter is the difference between the sums of the pixels within two rectangular regions 301-302. The regions have the same size and shape and are horizontally, see Figure 3A, or vertically, see Figure 3B, adjacent. A three-rectangle filter computes the sum within two outside rectangles 303 subtracted from twice the sum in a center rectangle 304, see Figure 3C. Finally a four-rectangle filter computes the difference between diagonal pairs of rectangles 305-306, see Figure 3D.
It should be noted that tens of thousands other simple
configurations of rectangle filters can be used. The filters canbe ofvarious sizes tomatchthe sizes of the detectionwindows . For two rectangle filters, the sum of intensities of the pixels within the unshaded rectangle are subtracted from the sum of the intensity of the pixels in the shaded rectangle. For three-rectangle filters, the sum of pixels in unshaded rectangle is multiplied by two to account for twice as many shaded pixels, and so forth. Other combinatorial functions can also be used with the filters according to the invention. We prefer simple operations for our filtersbecause they are very fast to evaluate, when compared with more complex filters of the prior art.
We also use a rectangle filter that has its internal components arranged diagonally. Figures 4A and 4C show variations of the rectangle filter that places the filters along a diagonal in the detection window 410. These diagonal filters 401-402 provide improved accuracy over the three types of filters described above for detecting non-upright faces and non-frontal faces. As shown in Figures 4B and 4D, the diagonal filters 401-402 are four overlapping rectangles 403-406, that combine to yield the blocky diagonal regions 408-409. These filters operate in the same way as the rectangle filters in Figure 3. The sum of the pixels in the shaded region 408 is subtracted from the sum of thepixels in shaded region 409. Diagonal filters are sensitive to objects at various orientations. The angle of the diagonal canbe controlledby the aspect ratios of the component rectangles
within the filter. Depending on their design, these rectangle filters can be evaluated extremely rapidly at various scales, orientations, and aspect ratios to measure region averages.
It is to be understood that various other adaptations and modifications may be made within the spirit and scope of the invention. Therefore, it is the object of the appended claims to cover all such variations and modifications as come within the true spirit and scope of the invention.