Patent Application
20080007747
Further aspects and advantages of the present invention will become apparent upon reading the following detailed description taken in conjunction with the accompanying drawings, in which:
Aspects of the invention are more specifically set forth in the following description with reference to the appended figures. Although the detailed embodiments described below relate to face recognition or verification, principles of the present invention described herein may also be applied to different object types appearing in digital images.
An input image is provided which may be a digital image obtained from any known image acquisition device, including, for example, a digital camera, a scanner, etc. The input image may also be an image created through any known synthetic techniques, such as computer generated animation, or may be a combination of digital data which is acquired via a sensor and synthetically generated. The input image may first undergo geometric normalization 105, shown in
Generally speaking, model 115 may provide any additional information which cannot be determined from the input image itself. Model 115 may be based on any model one of ordinary skill in the art could utilize in order to provide additional information to model-based anisotropic diffusion 110 better remove noise from the input image while minimizing the impact on image features corresponding to real structure.
Model 115 may be based on sensor noise characteristics, whereby if high sensor noise is detected in a region of the model, more filtering may be performed. Model 115 may be based on known geometric-based rules of objects within an image. For example, if an image of a face had dark and light nose halves, a geometric rule could be implemented into model 115 that dictates that noses of people do not have facial features divided in this way, and can this information can be supplied to model based anisotropic diffusion 110 to adjust filtering in this region accordingly. Model 115 may be based upon prior knowledge of controlled lighting which was in place when the image was acquired, and this knowledge may be used to determine the state of edges of the subject in the image. For example, in high illumination areas, one could utilize the knowledge that edges not due to structure in these areas would be minimized in these areas, and model based anisotropic diffusion 110 would vary the filter accordingly.
Model 115 may be based upon texture, and the likelihood that some textures have fewer edges than other textures. This information can be used to alter the filtering parameters in model based anisotropic diffusion 115. Proximity information may also be used to tell model based anisotropic diffusion 110 that various edges may not occur within a certain distance of each other. Markovian models may be used as a statistical approach to provide proximity information. Other models known in the art may also be used to provide such proximity information. Model 115 may also use a priori knowledge of periodic information to assist model based anisotropic diffusion 110. For example, if artifacts such as moiré occur in a particular application, the pattern may be modeled using techniques known in the art to provide information regarding edges to model based anisotropic diffusion 110 which are due to these periodic patterns, so they can be filtered accordingly. One of ordinary skill in the art would appreciate that other known models may be used in model 115 in various other embodiments consistent with the invention.
In one embodiment, model 115 can provide object-based information, and more specifically, information regarding edges within a representative object, which may include the location and likelihood of real edges within the representative object. As used herein, the term “real edges” may be defined as localized contrast variations (i.e. an edge) within an image which solely result from features associated with an object. The real edges typically may not be caused by other effects external to the object, such as environmental phenomena or sensor artifacts. For example, as described in more detail below, if the representative object in model 115 is a face, real edges indicated in the model may be the result of the structural variations in the features naturally occurring in a face, such as, for example, the eyes, nose, mouth, etc. Other representative objects may be generated depending upon what artifacts need to be removed in the input image. For example, predictable structural variations in an imaging device, such as, for example, lens or sensor imperfections, could be characterized in model 115 to assist the noise removal/image enhancement process.
Model 115 may be represented using a variety of different methods. One representation may include a multi-dimensional mathematical function which indicates the probability of an edge as a function of pixel position within the input image. The mathematical function could be determined using regression or other modeling techniques. Model 115 may also be represented by a two-dimensional dataset, having a structure like an image or a surface, where pixel indices in the horizontal and vertical directions represent location, and pixel values represent the probability of a real edge. The values pixel values may take on values between 0 and 1. Details regarding one embodiment for creating a model are presented in further detail below in the description of
Model based anisotropic diffusion 110 may perform the well known anisotropic diffusion process while utilizing real edge information supplied by model 115. While embodiments herein so far have described using anisotropic diffusion, other embodiments of the invention may contemplate other types of diffusion processes, which are known in the art, that could benefit from the information supplied by model 115.
Like standard diffusion algorithms, model based anisotropic diffusion (MBAD) 110 may iteratively perform noise reduction filtering over successive time periods, and use gradient information to determine whether or not an edge exists underneath a filter kernel for a given iteration. However, to improve the edge detection process, MBAD 110 can utilize information from model 115 to determine if an edge underling the filter kernel is a real edge. This information may be utilized during each iteration in the diffusion process, therefore this embodiment of MBAD 110 can modify the internal operations of the algorithm. These modifications can be implemented in a variety of ways. For example, model information 115 may be used to determine whether or not to apply filtering. In another example, model 115 information may be used to alter the filtering parameters, which is described in more detail below in
The filter parameters may be selected in a variety of different ways in step 215. In one embodiment, the actual size of the filter kernel could be varied. If the probability of an edge is indicated as high, the size of the filter kernel could be reduced, thus reducing the noise filtering effects. If the probability of a real edge is low, the size of the filter kernel could be increased to better reduce noise. In another embodiment, the values of the filter coefficients themselves may be changed based upon the value of the probability of a real edge. These values could be determined by a look-up table based upon real-edge probabilities, or they could be determined by a mathematical function known to one of ordinary skill in the art. In a simple embodiment, one may adjust the filter parameters so no filtering is performed when the probability of a real edge exceeds a threshold value. Once the filter parameters are determined, the image may be filtered in step 220 using the selected parameters. The filtering may be standard convolutional filtering, or any other filtering known to one of ordinary skill in the art.
The real edge information provided by model 115 may be solely used in selecting filter parameters, or this information may be combined with the gradient edge information typically provided by the anisotropic diffusion process. How these two types of information may be combined may be based upon by the level of confidence in the model of the representative object itself, and/or information regarding the conditions from which the input image was collected.
An input image may first undergo an optional geometric normalization step 105, which may be the same process described above in image processing method 100 shown in
Model 115 supplies real edge information to model application 310. This information may be combined with the diffusion image and the input image to improve the filtering process. In one embodiment, the diffusion image and the input image may be combined using a simple linear combination, wherein values from model 115 provide weights. The combination may be mathematically described by following equation:
O(x,y)=I(x,y)[1−M(x,y)]+D(x,y)M(x,y)
where
So for example, in areas of the input image where the edge probability is low, M(x,y) may take on values close to zero. In these areas, output image O(x,y) will be similar to the input image I(x,y).
Mass storage 520 can also store program instructions and digital data, and communicate to processor 510 over data bus 540. Processing system can provide and receive other information through I/O interface 525 and network interface 527, to provide information to users on display 530, and receive user commands and/or data through user I/O interface 535.
Although detailed embodiments and implementations of the present invention have been described above, it should be apparent that various modifications are possible without departing from the spirit and scope of the present invention.
