Image analysis – Pattern recognition
Reexamination Certificate
2000-10-30
2004-05-18
Ahmed, Samir (Department: 2621)
Image analysis
Pattern recognition
C382S190000, C382S224000, C382S225000, C382S226000, C382S227000, C382S228000
Reexamination Certificate
active
06738513
ABSTRACT:
The present invention relates to a method for fractal-darwinian object generation and in particular to a method wherein objects or structures to be examined are acquired in complex contexts, and corresponding objects or structures are generated. Herein both the overall method and the objects as well as their linkages may be considered to be fractal-darwinian.
Darwinian methods, or genetic or evolutionary methods, respectively, are used if the number of possibilities is so high as to preclude calculating all of them. In the presently described method, such genetic or evolutionary optimisation methods are combined with a “multi-scale”, fractal manner of proceeding. The interconnection between the “small one and the “large one” is thus dynamically taken into account. The term “fractal” is borrowed from Mandelbrot's fractal geometry while, however, having a different connotation hereinafter. As with Mandelbrot, a sub-object or a substructure resembles the object or structure it belongs to: i.e., the branch resembles the tree. In the method described hereinbelow, the objects might resemble their sub-objects, but generally they will not. Rather, the treatment and description of the objects will be similar for sub-objects and objects.
A like fractal-darwinian object generation is particularly utilised for image recognition or image generation, however may also be employed for speech recognition or speech generation, and in general for the recognition and generation of structures of all kinds presenting a particularly complex overall structure.
Conventional image recognition methods are based, for example, on direct comparison between a structure to be examined (e.g., image or an object within the image) and a structure already stored, which is compared directly with the structure to be examined. What is, however, a drawback in the like image recognition systems is the fact that owing to the direct comparison, on the one hand the structure indeed must be present in a very similar form, or any possible alterations and distortions must equally be tried, and on the other hand the objects frequently can be recognised not due to their properties, but partly or in extreme cases even primarily through their context, which means the close or remote surroundings. In addition, the accuracy and reliability of image recognition or image generation is limited inasmuch as an enormously high demand for storage capacity is required for the comparative structures.
The invention is therefore based on the object of furnishing a method for fractal-darwinian object generation whereby objects in a complex interrelation or having a complex structure which are to be examined may be recognised and imitated at high accurracy reliability.
In accordance with the invention, this object is attained by the measures indicated in claim
1
.
Herein a fractal-hierarchical object library including predetermined objects stored therein and presenting associated rules of property, context and alteration is used. The property rules describe the object, the context rules describe the relations between the objects (hierarchical and non-hierarchical relations), and the alteration rules describe the changes of the objects due to their properties and/or relations. Initially basic objects having subordinate and superordinate objects are hierarchically formed with relative formal algorithms, or generated from the complex structure of hierarchical basic structures to be examined, respectively. These basic objects are subsequently compared with the objects of the fractal object library, with a local classification likelihood being associated to each basic object by means of the property rule of the corresponding object in the object library. Thereupon further context rules are applied to the respective objects, wherefrom global or fractal classification likelihoods result. Finally, the fractal classification likelihoods of the basic objects are optimised in accordance with alteration rules. This method is performed repeatedly in order to improve the classification likelihoods.
In particular, in the step of forming the basic objects a method for iterative segmentation of basic elements may be utilised by taking into account relative formal homogeneity criteria (e.g., similar color, color fluctuation, etc.), resulting in particularly rapid object generation.
Moreover the fractal object library may present alteration rules encompassing prescriptions for an alteration—with alteration in the present instance denoting a change e.g., the propagation of the pattern or object (e.g., owing to more or less purposeful changes of shape)—or a comminution of an object, with generation of the subordinate and superordinate objects taking place in accordance with the growth rules and the amount of the respective classification likelihood.
Preferably objects are generated by fusing or founding, wherein the substructure is either dissolved or maintained. The generation of subordinate objects from basic objects or superordinate objects takes place by means of comminution or by means of known segmentation methods.
Preferably unknown and constantly recurring objects obtained, for example, during segmentation but not possessing a corresponding object in the fractal object library, may be included in the fractal object library together with associated rules of property, context and/or alteration. Hereby the so-called “wealth of experience” for the method may be expanded at will.
The property rules in the object library, which pertain to a particular object, set the properties of a particular object. For the case of image recognition or image generation, for example the shape, density and/or color of a particular object determine its properties. In speech recognition or speech generation, these properties may also be set in the form of pitch, tone, rhythm, etc. The context rules may consist of internal and external context rules, with the internal context rules determining a relation between equivalent objects within a same hierarchy. The objects may originate both from a common superordinate object and from different superordinate objects. The external context rules usually determine a relation between the subordinate and superordinate objects. Upon use of several hierarchical object structures among each other, the external context rules may, however, also determine a relation of objects within the hierarchical object structures existing in parallel.
Preferably the repeated execution of the method steps is terminated when the classification likelihood for the entire object to be examined has exceeded a particular threshold. As an alternative, the repeated execution of the method may also be terminated when a substantially stable overall object has been generated.
REFERENCES:
patent: 5461698 (1995-10-01), Schwanke et al.
patent: 5533148 (1996-07-01), Sayah et al.
patent: 6018728 (2000-01-01), Spence et al.
patent: 6397200 (2002-05-01), Lynch et al.
You, “Texture class assignment in texscale: an evaluation study”, IEEE, Acoustics, Speech, and Signal Processing, 1994; pp.: V/393-V/396 vol. 5.*
Danie, “Using Discriminant Eigenfeatures for Image Retrieval”, IEEE, Transactions on Pattern Analysis and Machine Intelligence, 1996; vol. 18.*
Dong, “Adaptive image restoration using a rule-based knowledge system”, IEEE, System Engineering, 1989; pp.: 19-22.
Baatz Martin
Binnig Gerd
Eschenbacher Peter
Schmidt Günter
Voss Richard
Ahmed Samir
Blakely & Sokoloff, Taylor & Zafman
Definiens AG
Le Brian
LandOfFree
Fractal production of darwinian objects does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Fractal production of darwinian objects, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Fractal production of darwinian objects will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3204579