Image analysis – Pattern recognition – Context analysis or word recognition
Reexamination Certificate
1999-04-17
2004-05-25
Johnson, Timothy M. (Department: 2625)
Image analysis
Pattern recognition
Context analysis or word recognition
C382S113000, C382S169000, C382S190000, C382S276000, C382S305000, C706S045000, C707S793000
Reexamination Certificate
active
06741744
ABSTRACT:
FIELD OF THE INVENTION
The present invention pertains to a method of object recognition and, more particularly, to a method of object recognition that uses a human-like language, based on the vocabularies used in photointerpretation, to write solution algorithms.
BACKGROUND OF THE INVENTION
In the art of object recognition, one usually extracts an object with image and/or map data by using one of three major methods: (1) a manual method in which an analyst extracts an object by using the human visual system, (2) an automated method in which the analyst relies totally on a machine system to perform the task, and (3) an interactive method in which the analyst determines a final decision, while a machine plays an assistant role.
Using a manual mode, an analyst does not need to employ a computer to do anything except to display an image.
In employing an automated system, once the data is entered into the machine, a machine system extracts the intended object. The analyst is merely a receiver of the data-processing results. In the event that the analyst is dissatisfied with the performance of the machine system, necessary changes can be made to the solution algorithms. In this automated mode, the analyst still has nothing to do with either the machine or the means by which objects are extracted.
In a conventional, interactive mode of information processing, the level of interaction between an analyst and a machine system can vary greatly. The least amount of interaction occurs when a machine system provides a set of solutions to the analyst, and the analyst selects or rejects one or more of the proffered solutions. On the other hand, the analyst can intensively interact with the machine by employing the following: (1) pre-processing image data by using a set of functions provided by the machine systems; (2) analyze the content of the scene by using a set of functions provided by the machine systems; (3) by utilizing the information provided in the aforementioned options, performing a set of object extraction options; and (4) evaluating each result and then selecting or rejecting a result.
In a conventional system utilizing intense interaction, an analyst is still either a mere operator or, at best, an effective, efficient user of the machine system. In other words, under these conditions, no matter how good the analyst is in the extraction or recognition of an object, conceptualized “algorithms” cannot be converted into a computer-workable program.
Conventional feature and object extraction for mapping purposes is based on high resolution panchromatic images supplemented by color imagery. A feature boundary, such as the contact zone between a forested area and a cultivated field, is determined by using standard photo-interpretation principles, such as a Tone principle, a Texture principle, a Size principle, a Shape principle, and so on, based on one single image. The use of multiple images, such as a system with three graytone images representing the near infrared spectrum, the red spectrum, and the green spectrum, in determining an object boundary can be very confusing and time-consuming. Therefore, to be helpful, these multispectral imagery data sets must be converted to a single-band scene serving as a base image map (IM) for manually-based feature extraction.
In the past 30 years or more, image processing and pattern recognition have been centered on extracting objects using simple and complex algorithms within an image of appropriate dimensions, such as 128×128, 256×256, 512×152 and 1024×1024 pixels. It is extremely rare for a complex algorithm to extract an object from a scene larger than 2048×2048 pixels, in view of the fact that historically even a workstation has a limited memory capacity to handle large images.
From the above discussion, it is clear that there exists a gap in the concept of scale in a physical space, and a gap in formation processing between the mapping community and pattern recognition scientists. In essence, cartographers deal with space in degrees of longitude and latitude, whereas image processing scientists think in terms of objects in a scene of 512×512 pixels. Among other objects of this invention, this conceptual and information processing gap is to be bridged.
The present invention is an innovative object-recognition system that divides objects into two broad categories, viz., wherein an analyst can articulate, after examining the scene content; how he or she would extract an object, and, secondly, wherein an analyst cannot articulate how to discriminate an object against other competing object descriptors, after examining the scene or a set of object descriptors (e.g., a spectral signature or a boundary contour).
In the first case, where an analyst is able to articulate the extraction of the objects, the proposed solution is to employ a pseudo-human language, including, but not limited to, pseudo-English as a programming language. The analyst can communicate with a machine system by using this pseudo-human language, and then inform the machine how he or she would extract a candidate object without having to rely on a “third-party” programmer.
In the second case, where an analyst is unable to articulate the extraction of an object, the proposed solution is to use an appropriate matcher with a matching library to extract the candidate object, and then pass it over to processors employed in the first-category sphere. Once an extracted object is passed over to the first environment, this object becomes describable by using the proposed pseudo-human language. Thus, it can be combined with other “existing objects” to extract still further objects. The final result, then, is the extraction of a set of complex objects or compound objects.
In the past 50 years, photointerpreters have been taught to use the principles governing the following aspects in recognizing an object: (1) tone or spectrum principles; (2) texture (spatial variation of tones) (3) size; (4) shape; (5) shadow (detection of vertical objects); (6) pattern (geometry and density) (7) associated features (contextual information); and (8) stereoscopic characteristics (height), if available.
From these principles based on the human visual system, an object can exist in a number of forms, as shown below in Table I.
TABLE 1
Object Form and Its corresponding Major Extraction
Principle Used By A Photo-interpreter or Image Analyst
Object Existing as:
Major Extraction Principles:
a. One pixel (sub-pixel) object
Tone or multi-spectral data
b. One-pixel region
Tone and size
c. Multiple, one-pixel object
Tone, pattern
d. One multiple-pixel region
Tone, texture, size, shape
e. Multiple contiguous regions
Tone, texture, size, shape, associated
features
f. Multiple discrete regions
Tone, size, texture, pattern, associated
features
g. Association with Others
Associated features, tone, size,
texture, etc.
From Table I data, it can be inferred that a spectral-matching-alone system can extract only one of the seven object types, i.e., (a). A shape-alone system can extract only two of the seven object types, i.e., (b) and (d).
The proposed system of this invention is intended for extracting all seven types of objects by using image and map data, such as synthetic aperture radar (SAR), and multi-spectral and other types of sensory data, with the assumption that appropriately matching libraries are available. For example, such libraries are readily available: (1) hyperspectral library of various material types; (2) ground vehicle library for FLIR (forward-looking infrared) applications; and (3) ground vehicle library for LADAR (laser radar) applications.
Using these libraries, the method of this invention first extracts single-pixel and single-region-based objects, and then “glues” them together to form multi-object-based object complexes.
Table II below illustrates this two-stage, object-extraction approach.
The uniqueness of this inventive method lies in using a pseudo-human language (such as a pseudo-English-based programming language), compatible with an interpreters&apo
Chawan Sheela
Johnson Timothy M.
Salzman & Levy
LandOfFree
Compiliable language for extracting objects from an 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 Compiliable language for extracting objects from an image..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Compiliable language for extracting objects from an image... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3185445