Feature extraction using pixel-level and object-level analysis

• US 7,933,451 B2
• Filed: 11/13/2006
• Issued: 04/26/2011
• Est. Priority Date: 11/23/2005
• Status: Active Grant

• Current Assignee: Leica Geosystems AG (Hexagon AB)
• Original Assignee: Leica Geosystems AG (Hexagon AB)
• Inventors: Kloer, Brian
• Primary Examiner(s): Ahmed; Samir A
• Assistant Examiner(s): Park; Edward
• Application Number:
  US11/559,237
• Publication Number:
  US 20070116365A1
• Time in Patent Office: 1,625 Days
• Field of Search: 382/190, 382/200, 382/206, 382/256, 382/259, 382/181, 382/223
• US Class Current: 382/190
• CPC Class Codes: G06F 18/211 Selection of the most significant subset of features
• CPC Class Codes: G06T 17/05 Geographic models
• CPC Class Codes: G06T 2207/10032 Satellite or aerial image; Remote sensing
• CPC Class Codes: G06T 2207/30184 Infrastructure
• CPC Class Codes: G06T 7/11 Region-based segmentation
• CPC Class Codes: G06T 7/136 involving thresholding
• CPC Class Codes: G06T 7/143 involving probabilistic approaches, e.g. Markov random field [MRF] modelling
• CPC Class Codes: G06T 7/49 based on structural texture description, e.g. using primitives or placement rules
• CPC Class Codes: G06V 10/255 Detecting or recognising potential candidate objects based on visual cues, e.g. shapes

Abstract

Image processing for extracting features in images. Pixel-level cue algorithms can be performed on raster images. The raster images can be converted to a vector layer. Object-level cue algorithms can be performed on the vector layer. The feature can be detected using a result of the pixel-level cues and using a result of the object-level cue algorithms performed. A computer-readable medium can include a first data field containing data representing pixel-level cues functioning to describe a pixel-level cue of the feature. The computer-readable medium can also include a second data field containing data representing object-level cues functioning to describe the object-level cues of the feature. Relation-level cue algorithms can be performed on the vector layers. The features can be detected using a result of any combination of the pixel-level cue algorithms, object-level cue algorithms, and/or relation-level cue algorithms.

Claims

1. In a graphical information system (GIS), a method for processing a digital image depicting one or more physical objects to locate a feature in the digital image, the method comprising the following:

by a processor, executing a pixel-level cue algorithm to identify an interesting area of a raster image depicting the one or more physical objects;

determining a pixel-level probability that the interesting area of the raster image identified is the feature using a result from the pixel-level cue algorithm;

comparing the pixel-level probability to a pixel-level cue threshold; and

if the pixel-level probability satisfies the pixel-level cue threshold;

converting at least a portion of the raster image to a vector layer by geometric modeling using points, lines, curves, and polygons to generate a representation of digital vector objects represented by the points, lines, curves and polygons;

executing an object-level cue algorithm on the vector layer to identify an interesting area of the vector layer;

determining an object-level probability that the interesting area of the vector layer is the feature using a result of the pixel-level cue algorithm; and

comparing the object-level probability to an object-level threshold.

2. The method according to claim 1, further comprising:

if the object-level probability meets the object-level threshold;

identifying the interesting area of the vector layer as the feature.

3. The method according to claim 1, further comprising:

if the object-level probability meets the object-level threshold;

executing a relation-level cue algorithm on the vector layer to identify an interesting area of the vector layer;

determining a relation-level probability for the interesting area of the vector layer using a result of the relation-level cue algorithm;

comparing the relation-level probability to a relation-level threshold; and

if the relation-level probability satisfies the relation-level threshold;

identify the interesting area of the vector layer as the feature.

4. The method according to claim 3, wherein the relation-level cue comprises one of the members of a set comprising:

a pattern cue; and

an association cue.

5. The method according to claim 1, wherein a training cue metric is input by a user.

6. The method according to claim 5, wherein the cue metric is at least one of:

a training pixel; and

a training vector object.

7. The method according to claim 1, wherein:

the pixel-level cue comprises at least one of;

a spectral cue;

a texture cue;

a vegetative cue;

a site cue;

a slope cue;

an aspect cue;

a landcover cue;

a soil cue;

a temperature cue; and

a user defined pixel-level cue; and

the object-level cue comprises at least one of;

a shape cue;

a size cue;

a orientation cue;

a right angle cue;

a template matching cue;

an orthogonality cue;

a perimeter cue;

a length cue;

a width cue;

a boundary complexity cue;

a rectangular fit cue;

an elliptical fit cue;

a compactness cue;

a symmetry cue;

a direction cue;

a background cue;

a spectral information of surrounding pixels cue;

a textural information of surrounding pixels cue;

an extent cue;

an area cue;

a shadow cue;

a background cue; and

a zonal cue.

8. The method according to claim 1, wherein the probabilities are calculated using one of:

a Bayesian Network;

Classification and Regression Trees;

or a Neural Network.

9. The method according to claim 1, further comprising:

extracting the feature; and

generating geometric object data representing the feature.

10. The method according to claim 1, further comprising:

based on the results of the pixel-level cue algorithm, generating a pixel probability layer in which each pixel'"'"'s value represents a probability that it is the feature of interest;

converting the pixel probability layer into a raster object layer which contains pixels that are grouped as raster objects; and

for each raster object calculating a mean probability of all of the probability pixels corresponding to each raster object.

11. The method according to claim 10, wherein at least a portion of the raster object layer is converted to a digital vector object layer by geometric modeling, wherein the geometric modeling uses geometric primitives including at least one of points, lines, curves and polygons to generate a representation of one or more digital vector layer objects represented by points, lines, curves and polygons.

12. The method according to claim 11, wherein each digital vector layer object is associated with the mean probability calculated from the raster object from which the vector object was converted.

13. A non-transitory computer-readable medium including computer-executable instructions for performing the method of claim 1.

14. In a graphical information system (GIS), a method for processing a digital image depicting one or more physical objects to locate a feature in the digital image, the method comprising the following:

by a processor, performing a pixel-level cue algorithm on a digital raster image, the raster image depicting the one or more physical objects by a set of pixels;

based on the results of the pixel-level cue algorithm, generating a pixel probability layer in which each pixel'"'"'s value represents a probability that it is the feature of interest;

converting the pixel probability layer into a raster object layer which contains pixels that are grouped as raster objects;

for each raster object, calculating a mean probability of all of the probability pixels corresponding to each raster object;

converting at least a portion of the raster object layer to a digital vector layer by geometric modeling, wherein geometric modeling uses geometric primitive including points, lines, curves and polygons to generate a representation of digital vector objects represented by the points, lines, curves and polygons, and wherein each digital vector layer object is associated with the mean probability calculated from the raster object from which the vector object was converted;

performing an object-level cue algorithm on the digital vector layer representation of the one or more digital vector layer objects; and

identifying a feature using a result of the pixel-level cue algorithm and a result of the object-level cue algorithm performed.