Loading patent…
AssigneeCommissariat a l'Energie Atomique
InventorsCorinne Thomas, Hassane Essafi
Estimating the term…
An initial image is subjected to a plurality of segmentations into uniform components by using various criteria to identify and isolate several series of objects constituted by segmented thumbnail images associated with geometrical descriptors for boxes that cover said objects, the results of the segmentation methods are fused, numerical values are used to characterize each object thumbnail image that results from fusion, and each of the thumbnail images is compared with indexed thumbnail images in an indexed database containing a set of textual descriptions for everyday elementary objects. The list of textual descriptions of everyday elementary objects matching the object thumbnail images that result from the segmentation and fusion operations is provided together with pertinence levels.
1/ A method of automatically creating an image database that is capable of being interrogated by its semantic content, the database being created from initial images that are not indexed,
2/ A method according to claim 1, characterized in that after step m) of determining the geometrical locations of the objects relative to one another and of textually characterizing the spatial relationships between the various objects, performing a step o) of establishing whether an initial image belongs to a theme on the basis of dependency relationships between the various objects and a comparison with the thematic information dictionary.
3/ A method according to claim 1, characterized in that after step m) of determining the geometrical locations of the objects relative to one another and of textually characterizing the spatial relationships of the various objects, performing a step p) in which the textual information for an entire indexed image is brought together in association with the vade-mecum.
4/ A method according to claim 1, characterized in that during step e) of classification in the image database, special features of the initial image under analysis are characterized relative to each of the most similar previously-indexed images.
5/ A method according to claim 1, characterized in that step h) of fusing the results of the representation methods consists in ordering the set of objects that results from the segmentation in order of decreasing area so as to obtain a list, and then so long as there remain unvisited objects in the list, in taking the first free element of the list, in checking to see whether said elements overlap any of the following elements in the list that are free, and if there is an overlap between objects derived from different segmentation methods, in constructing a first graph and in marking said elements as being not free, so as to obtain a series of first graphs, in performing the same operations again using the same objects that result from segmentations ordered in order of increasing area, so as to obtain a series of second graphs, and in combining the first and second graphs so as to fuse the results of the segmentation methods.
6/ A method according to claim 1, characterized in that during step g) of segmenting locally analyzed images into uniform components, by using a plurality of different criteria, the following are performed:
7/ A method according to claim 6, characterized in that during segmentation step g), the detection of texture characteristics via filtering equivalent to a bank of highpass filters and lowpass filters comprises a wavelet transformation process performed firstly on rows and then on columns of the image with on each occasion a lowpass filter characterized by the formula:
8/ A method according to claim 7, characterized in that after each operation of filtering by means of a highpass filter and a lowpass filter, the number of images obtained is divided by 2n so as to reduce the amount of data to be processed, the integer n increasing with increasing resolution and size of the image to be processed.
9/ A method according to claim 6, characterized in that the wavelet transformation process is reiterated several times over n stages, and in that after a smoothing stage, the size of the resulting 3D attribute matrix is reduced in depth by retaining only the [2n+(n−1)] images that contain the most detail.
10/ A method according to claim 7, characterized in that a process is also performed of seeking the distribution of the attribute matrix that is best in the sense of maximizing energy.
11/ A method according to claim 6, characterized in that during segmentation step g), chromatic and achromatic detection comprises a step of converting the image to the hue saturation value (HSV) model, an achromatic propagation step to eliminate the background, an opening step performed during a first iteration to eliminate noise and reconnect regions, and in the event of non-convergence at the end of this first iteration, a closure step performed during a second iteration.
12/ A method according to any one of claims 6 to 11, characterized in that during segmentation step g), the detection of outlines using an optimum filter comprises a step of converting the image to red-green-blue (RGB) in a single luminance plane, a step of averaging in order to eliminate noise associated with acquisition, a step of computing horizontal and vertical gradients, a step of summing the two gradient images in order to obtain a single image, a step of computing the minimum and the maximum of the resulting image in order to determine high and low thresholding values, a hysteresis thresholding step relative to said high and low threshold values as previously determined, a step of performing morphological opening, and a step of performing multiple morphological closures.