Browse Prior Art Database

Adaptive automatic determination of picture orientation using Data Mining methodology Disclosure Number: IPCOM000016223D
Original Publication Date: 2002-Sep-01
Included in the Prior Art Database: 2003-Jun-21
Document File: 4 page(s) / 54K

Publishing Venue




This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 50% of the total text.

Page 1 of 4

  Adaptive automatic determination of picture orientation using Data Mining methodology


Scenario and State of the art

User has used a digital camera picture format produced by the camera is rectangular with different length of x-axis and y-axis stored pictures can have "portrait" orientation or "landscape" orientation, depending on the orientation the camera at snapshot time at download time, a default orientation is chosen by the system, e.g. landscape for each picture, a human user looks at the pictures and determines whether the default orientation is correct, if the default orientation is not correct for a certain picture, the human has to trigger a rotate operation by 90 degrees to toggle "portrait" and "landscape" orientation.


This manual interaction degrades the comfort of the overall system (and makes it look stupid). Automatic determination of the orientation is not trivial since the intended orientation is context (scenario) dependent.


Use Data Mining Techniques to adjust to the context: learn a Data Mining model how to determine the intended orientation from a training set of given pictures; Then determine the orientation to new pictures by applying that data mining model.


Perform the following steps

1.Provide a training set of pictures ST where the orientation is the desired one. 2.Provide a candidate set of pictures SC for which you want to adjust the orientation. 3.For each picture of STU SC, use an image analysis tool to extract a set of features; for example, the IBM software QBIC (Query By Image Content) provides mechanisms to extract features like texture, and colors (color layout, color histogram, avg color). Store the data generated in 3 together with orientation info in a table T;


5.For the records in T corresponding to set STrun a Data Mining prediction method in Training mode which generates model M.

6.For the records in T corresponding to set SC run the Data Mining prediction method in Application mode using model M to predict the orientation. 7.For each element from SC check whether the orientation is identical with the


Page 2 of 4

predicted orientation; depending on the confidence, propose to change orientation or even change the orientation without a prompt.

Remarks: ad 3: Details for QBIC (excerpt from, Dec 2001): "Currently QBIC supports several basic image similarity measures such as: average color, color histogram, color layout, and texture. They are all computed over the whole image or previously masked area (good when main objects are within large background area). Average color is recommended for images with unformed color. Color histogram is recommended when you are interested in overall color content, irrespective where the colors are in the image (wallpapers, ties etc.). Color layout matches on colors and their position and gives more precise matching (good for stock photo search). Texture is computed as d...