Image analysis – Image segmentation
Reexamination Certificate
2000-12-01
2004-09-21
Tran, Phuoc (Department: 2621)
Image analysis
Image segmentation
C382S199000
Reexamination Certificate
active
06795577
ABSTRACT:
The present invention concerns in general terms the analysis of a digital signal and proposes for this purpose a device and method for analysing a digital signal by decomposition into a plurality of resolution levels, and segmentation.
The purpose of the analysis is to provide a hierarchical segmentation of the signal, thus make it possible to access the objects or regions present in an image, at several resolution levels, with several possible levels of detail. Access to the objects of an image can be used for different purposes:
selective coding of the objects of the image, granting a higher coding quality to the “important” objects in the image,
progressive transmission of the date of the image, with transmission of the more important objects before the others,
extraction of a particular objects from the image, with a view to its manipulation, transmission, coding, and storage.
The present invention is more particularly applicable to the analysis of a digital signal. Hereinafter, the concern will more particularly be with the analysis of digital images or video sequences. A video sequence is defined as a succession of digital images.
There exist several known ways of effecting the decomposition of a signal on several resolution levels; it is for example possible to use Gaussian/Laplacian pyramids, or to decompose the signal into frequency sub-bands at several resolution levels.
The remainder of this description will be concerned with the second case, but it is important to note that the present invention applies to all known multi-resolution decompositions.
In the particular case of a decomposition into frequency sub-bands, the decomposition consists of creating, from the digital signal, a set of sub-bands each containing a limited frequency spectrum. The sub-bands can be of difference resolutions, the resolution of a sub-band being the number of samples per unit length used for representing this sub-band. In the case of an image digital signal, a frequency sub-band of this signal can be considered to be an image, that is to say a bi-dimensional array of digital values.
The decomposition of a signal into frequency sub-bands makes it possible to decorrelate the signals so as to eliminate the redundancy existing in the digital image prior to the compression proper. The sub-bands can then be compressed more effectively than the original signal. Moreover, the low sub-band of such a decomposition is a faithful reproduction, at a lower resolution, of the original image. It is therefore particularly well suited to segmentation.
The segmentation of a digital image will make it possible to effect a partitioning of the image into homogeneous regions which do not overlap in this context, the image is considered to consist of objects with two dimensions. The segmentation is a low-level process whose purpose is to effect a partitioning of the image into a certain number of sub-elements called regions. The partitioning is such that the regions are disconnected and their joining constitutes the image. The regions correspond or do not correspond to objects in the image, the term objects referring to information of a semantic nature. Very often, however, an object corresponds to a region or set of regions Each region can be represented by information representing its shape, colour or texture. The homogeneity of the region of course depends on a particular criteria of ho homogeneity; proximity of the average values or preservation of the contrast or colour, for ample.
Object means an entity of the image corresponding to a semantic unit, for example the face of a person. An object can consist of one or more regions contained in the image. Hereinafter the term object or region will be used indifferently.
Conventionally, the segmentation of the digital image is effected on a single resolution level, which is the resolution of the image itself. Conventionally, the segmentation methods include a first step known as marking, that is to say the interior of the regions housing a local homogeneity is extracted from the image Next a decision stop precisely defines the contours of the areas containing homogeneous data. At the end of this step, each pixel of the image is associated with a label identifying the region to which it belongs. The set of all of the labels of all the pixels is called a segmentation map,
This type of segmentation makes it possible to obtain a relatively effective segmentation of the image but has the drawbacks of being slow and not very robust and presenting all the objects at the same resolution.
This is the case for example with the so called MPEG4 standard (from the English “Motion Picture Expert Group”), for which an ISO/IEC standard is currently being produced, in the MPEG4 coder, and more particularly in the case of the coding of fixed images, the decomposition of the image into frequency sub-bands is used conjointly with a segmentation of the image. A step prior to coding (not standardised) is responsible for isolating the objects of the image (video objects) and representing each of the these objects by a mask. In the case of a binary mask, the spatial support of the mask has the same size as the original image and a point on the mask at the value 1 (or respectively 0) indicates that the pixel at the same position in the image belongs to the object (or respectively is outside the object).
For each object, the mask is then transmitted to a shape decoder whilst the texture for each object is decomposed into sub-bands, and the sub-bands are then transmitted to a texture decoder.
This method has a certain number of drawbacks. This is because the object is accessible only at its highest resolution level there is no progressivity in segmentation. Moreover, the number of objects manipulated is a priori the same at all levels, whilst it may be more advantageous to have a number of objects increasing with the (spatial) resolution, that is to say a true conjoint scalability between the resolution and the number of objects.
The article “Multiresolution adaptative image segmentation based on global and local statistics” by Boukerroui, Basset and Baekurt, which appeared in IEEE international Conference on Image Processing, 24-28 Oct. 1999, vol. 1 pages 358 to 361, describes a hierarchical segmentation based on a multiresolution pyramid of an image, effected by discrete wavelet transform, known as DWT.
In addition, the article “Multiresolution image segmentation for region-based motion estimation and compensation” by Salgado, Garcia, Menendez and Rendon, which appeared in IEEE International Conference on image Processing, 24-28 Oct. 1999, vol. 2, pages 135 to 139, describes a hierarchical segmentation also based on a multiresolution pyramid of an image. A partitioning of the image effected at the lowest resolution level is projected onto the higher resolution levels.
However, none of these known methods provides access to the region or objects with different resolution levels, in a consistent and coherent manner. Coherent means here that an object at a given resolution level always descends from a single object with a lower resolution (parent), and gives rise to at least one object at the higher resolution level (child or children).
The present invention aims to remedy the drawbacks of the prior article by providing a method and device for the hierarchical segmentation of a digital signal which offers access to the regions or objects at different resolution levels, in a consistent and coherent manner.
In this context the invention concerns a method of analysing a set of data representing physical quantities, including the steps of,
decomposition of the set of data on a plurality of resolution levels,
segmentation of at least a sub-part of a given resolution level, into at least two homogeneous regions, said given resolution level not being the highest resolution level in the decomposition.
characterised in that it includes the steps of:
storage of information representing at least part of the result of the segmentation of the previous step,
segmentation of at least one sub-part of t
Canon Research Centre France S.A.
Fitzpatrick ,Cella, Harper & Scinto
Tran Phuoc
LandOfFree
Digital signal analysis, with hierarchical segmentation does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Digital signal analysis, with hierarchical segmentation, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Digital signal analysis, with hierarchical segmentation will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3214001