Image analysis – Pattern recognition – Classification
Reexamination Certificate
1999-03-16
2002-04-16
Johns, Andrew W. (Department: 2621)
Image analysis
Pattern recognition
Classification
Reexamination Certificate
active
06373984
ABSTRACT:
FIELD OF THE INVENTION
This invention relates to digital image processing, and more particularly, to pattern and object recognition in a digital image.
BACKGROUND OF THE INVENTION
Digital image processing systems that recognize patterns in images and detect objects within an image are well known. These systems use various conventional processing techniques for both of these tasks. For example, some of these systems are strictly mathematically based and use such algorithms as morphological processing, wavelet algorithms, convolutions, and the like while other systems use variations on neural net and fuzzy logic technologies. These systems have been used in a variety of image analysis environments and applications such as medical images obtained with ultrasound, x-ray, CAT scan, MRI and other known medical imaging systems to examine body tissue for the detection of abnormalities such as cancer. These systems have also been used in military image analysis of images generated by radar, sonar, infrared and other imaging systems to extract features such as terrain, infrastructure and the like and to detect specific targets such as tanks and other military vehicles, weaponry, camouflaged bunkers and the like. These systems have also been used in various commercial and scientific applications.
These previously known image analysis systems are extremely limited in the image environment in which they properly function. Additionally, the algorithms used by these systems to process images are brittle. That is, they are typically rendered ineffective by slight variations in image quality, image resolution, image noise level and many other factors. Consequently, these systems cannot be used for general pattern recognition and object detection across a wide range of applications, environments, or imaging modalities.
One of the most difficult image analysis environments is the ultrasound imaging of in vivo body tissues. The images of in vivo tissue generated by ultrasound equipment contain a high number of line variations and edge patterns, the detection of which is made difficult by the signal noise present in that imaging environment. Thus, image analysis based on a comparison of images of the same area of interest taken at different times, as typically done in systems analyzing ultrasound images, is inherently difficult.
With systems that use images to detect the presence of cancer, limitations arise from the knowledge base required by such systems and the computational resources available for image comparison. For example, neural networks may be used for image analysis systems; however, these systems require extensive training through the presentation of numerous images to the neural network. Because the images are presented to the system without intelligent insight as to the content of the images, the system tends to only learn statistically significant events and ignores image events less frequently encountered. For example, the submission of images of tissue having a particular abnormal cell structure indicative of some cancer may be learned by a neural network system to identify that structure and similar structures in tissue. However, a less statistically significant variant of that structure that may also indicate cancer would not be incorporated in the knowledge base of the system. Consequently, the system may not detect these variant structures which are also indicative of cancer. Another problem with previously known image analysis systems is the amount of resources required for image processing. The elements for processing images on a pixel by pixel basis can be significant as the number of pixels contained within an image increase. In some cases, the storage of multiple images so they may be later compared to one another can be prohibitive or require archival systems that increase the time required for image retrieval and processing.
What is needed is an image processing system and method that recognizes patterns and detects objects in images without requiring adaptation of the system to a particular application, environment, or image content.
What is needed is an image analysis system and method that maintains consistent image evaluation which is independent of variation in image modality, resolution, noise level, quality, and other image factors.
What is needed is an image analysis system and method for cancer detection that accurately analyzes in vivo images in a wide variety of applications.
What is needed is an image analysis system and method that simplifies the processing of images from the pixel by pixel comparisons performed by known systems.
What is needed is an image analysis system and method that incorporates within a knowledge base variant structures of an object or pattern that are not statistically significant.
What is needed is an image analysis system and method for feature extraction and target recognition in various military, commercial and scientific digital images.
SUMMARY OF THE INVENTION
The above limitations of previously known image processing systems are overcome by a system and method operating in accordance with the principles of the present invention. An image processing system made in accordance with the principles of the present invention includes a knowledge base of themes with the themes being coupled to knowledge elements by associative links, a synaptic link generator for generating synaptic links for pixels in a digital image, the synaptic link being used to identify a knowledge element for a corresponding pixel and a theme identifier for evaluating associative links for a plurality of knowledge elements corresponding to a theme to determine whether the corresponding theme is present in a digital image. This system may be used to analyze a digital image for patterns and objects in many different applications, environments, and imaging modalities.
In the system of the present invention, the synaptic link generator uses a data mask to select pixels from an image and generates a synaptic link from determinants generated for each pixel. Preferably, these determinants are generated from data for pixels lying in radials extending from the pixel under investigation. A plurality of determinants are preferably used to generate a synaptic link that identifies pattern information surrounding the pixel under investigation. These determinants encode pattern data in the vicinity of the target pixel so it may be more easily processed and recognized. A trainer then determines the identity of a knowledge element or knixel that corresponds to each pixel. This operation may be performed by having the trainer identify a window around an object within an image and identifying a knowledge element for that object so that each synaptic link generated for the pixels within the window is associated with the identified knowledge element. For example, an outline of a blood vessel may be defined within a digital image by a trainer and each synaptic link for each pixel within that window is associated with the knowledge element identified as a “blood vessel.” This type of learning is called area learning.
A target learning mode may also be used. In this learning mode, the synaptic links within a window are mapped to a knowledge element, then all portions of the image which the trainer recognizes as corresponding to the knowledge element being learned by the system are blocked out and the synaptic links for the remaining pixels of the image outside the active window and blocked out regions are set to a knowledge element identifier that indicates the knowledge element is unknown. This process breaks the synaptic links generated for “noise” pixels within the active knowledge element window to the knowledge element identifier being learned. By repopulating the blocked out areas with image data and processing them, a set of synaptic links for the knowledge element being learned is developed that identify the objects corresponding to that knowledge element from different perspectives.
During a learning phase, a synaptic link may be generated for association with a knowledge e
Gouge James
Gouge Sally
Azarian Seyed
Indigo Medical Incorporated
Johns Andrew W.
LandOfFree
System and method for detecting patterns or objects in a... 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 detecting patterns or objects in a..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and System and method for detecting patterns or objects in a... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2824538