Diferenças

Aqui você vê as diferenças entre duas revisões dessa página.

--- interimage:attributes_description [2010/06/17 22:41] – rsilva
+++ interimage:attributes_description [2010/06/23 17:09] – castejon
@@ Linha 1: / Linha 1: @@
 ======= Attributes Description =======
@@ Linha 28: / Linha 29: @@
   * **xGeoCenter** - x geo-coordinate of the object centroid.
   * **yGeoCenter** - y geo-coordinate of the object centroid.
-  * **membership** or **p** - confidence of the object with regard to its class.
+  * **membership** or **p** - confidence in the object with regard to its classification.
 ===== Shape Attributes =====
@@ Linha 56: / Linha 57: @@
  {{ :interimage:shapeindex.png }}
 Where P is the polygon perimeter and A is the area.
@@ Linha 62: / Linha 66: @@
 ===== Spectral Statistical Attributes =====
-  * **Amplitude** - Blablabla.
+  * **Amplitude** -
-  * **Brightness** - Blablabla.
+  * **Brightness** -
-  * **Correlation** - Blablabla.
+  * **Correlation** - Correlation is a similarity measures between two data sets under an absolute scale between [-1,1]. It is calculated as showed by the next formula:
+{{ interimage:att_correlation.gif }}
-  * **Covariance** - Blablabla.
+  * **Covariance** - The covariance value represents the similarity degree between two data sets showing how correlated they are. Higher data correlation leads to higher covariance values. The calculus is showed by the following formula where N is the number of image elements for one given area. X(i) are the element values for each given index "i".
+{{ interimage:att_covariance.gif }}
-  * **Entropy** - Blablabla.
+  * **Entropy** - This is a randomness statistical measure that can be used to describe some texture features. Higher data randomness leads to higher entropy values. The calculus is done as showed by the next formula, where n is the number of distinct image element values and p(xi) is the occurrence frequence associated to that pixel value.:
+{{ interimage:att_entropy.gif }}
-  * **MaxPixelValue** - The maximum pixel value found inside one image region.
+  * **MaxPixelValue** - The maximum pixel value found inside one region for the given image band/channel.
   * **Mean** - The mean of image elements numeric values X1,X2,...,Xn. It is expressed by:
 {{ interimage:att_mean.gif }}
-  * **MinPixelValue** - Blablabla.
-  * **Mode** - Blablabla.
+  * **MinPixelValue** - The minimum pixel value found inside one region for the given image band/channel.
-  * **Ratio** - Blablabla.
+  * **Mode** - Represents the most frequent value among a set of values. There are cases where mode value cannot exist and there are cases where its value it is not garanteed to be unique. Examples:
+    * 1,1,3,3,5,7,7,7,11,13 : Mode 7
+    * 3,5,8,11,13,18 : Mode value does not exists.
+    * 3,5,5,5,6,6,7,7,7,11,12 : Two mode values - 5 and 7 (bimodal).
+  * **Ratio** -
   * **StdDeviation** - The standart deviation represents the numerical data dispersion degree surrounding the mean. It is defined by:
 {{ interimage:att_stddev.gif }}
-  * **SumPixelsValues** - Blablabla.
+  * **SumPixelsValues** - Represents the sum of all elements values inside on area for one given image band/channel.
   * **Variance** - Like the standart deviation, the variance also represents the numerical data dispersion degree surrounding the mean but in the original data values scale. It is defined by:
 {{ interimage:att_variance.gif }}
-===== Spectral Texture Attributes =====
+===== Texture Attributes =====
+The texture attributes are based on the co-occurence gray scale matrix (GLCM) described by the following references:
+  * Textural Features for Image Classification - Robert M. Haralick, K. Shanmugam, Its'hak Dinstein. Systems, Man and Cybernetics, IEEE Transactions on In Systems, Man and Cybernetics, IEEE Transactions on, Vol. 3, No. 6. (1973), pp. 610-621.
+  * Computer and Robot Vision - Robert M. Haralick - Addison-Wesley Publishing Company.
+\\
+  * **Angular2ndMomentGLCM (a.k.a. EnergyGLCM)** - Returns the square sum of image points pairs occurrences under one pre-defined direction. The returned values range is between [0,1]. For those images without variations the value will be 1. The calculus is showed on the next formula where "i" and "j" are adjacent image points values under one pre-defined direction. p(i,j) is the probability of that co-ocurrence over the image.
+{{ interimage:att_angular2ndmomentglcm.gif }}
+  * **ContrastGLCM** - Returns a contrast intensity measure  between one image point and its neighborhood. For those images without variations the contrast value will be zero. The calculus is showed on the next formula where "i" and "j" are adjacent image points values under one pre-defined direction. p(i,j) is the probability of that co-ocurrence over the image.
+{{ interimage:att_contrastglcm.gif }}
+  * **DissimilarityGLCM** - Returns one intensity measure quite similar to contrast between one point and its neighborhood. But the difference it that this measure has linear increments. The calculus is showed on the next formula where "i" and "j" are adjacent image points values under one pre-defined direction. p(i,j) is the probability of that co-ocurrence over the image.
+{{ interimage:att_dissimilarityglcm.gif }}
+  * **EntropyGLCM** - Like the simple statistical entropy the GLCM entropy also is a statistical measure of image data randomness. The difference is that it uses frequencies of gray levels co-ocurrences instead of using point values frequencies. The co-ocurrences matrix is used and the calculus is showed by the next formula where "i" and "j" are adjacent image points values under one pre-defined direction. p(i,j) is the probability of that co-ocurrence over the image.
+{{ interimage:att_entropyglcm.gif }}
+  * **HomogeneityGLCM** - Returns a value representing the distance between the distribuition of co-ocurrence matrix elements and those diagonal elements. The returned values range is between [0,1]. For images with low values variation the returned value will be near to zero. The calculus is showed by the next formula where "i" and "j" are adjacent image points values under one pre-defined direction. p(i,j) is the probability of that co-ocurrence over the image.
+{{ interimage:att_homogeneityglcm.gif }}
+  * **MeanGLCM** - The GLCM mean value is expressed in function of the frequency of co-ocorrence of image elements related to their neighborhood under one pre-defined direction. The calculus is showed by the next formula where "i" and "j" are adjacent image points values under one pre-defined direction. p(i,j) is the probability of that co-ocurrence over the image.
+{{ interimage:att_meanglcm.gif }}
+  * **QuiSquareGLCM** - This metric can be understood as a form of energy normalization expressed in function of the linear dependency gray levels for image elements. The calculus is showed by the next formulas where "i" and "j" are adjacent image points values under one pre-defined direction. p(i,j) is the probability of that co-ocurrence over the image. Pj is the marginal probability for that co-ocurrence.
+{{ interimage:att_quisquareglcm_1.gif }}
+{{ interimage:att_quisquareglcm_2.gif }}
+{{ interimage:att_quisquareglcm_3.gif }}
+  * **StdDeviationGLCM** - The standart deviation is a measure that represents the values dispersion around a GLCM mean value. The GLCM standart deviation calcule differs from the simple standart deviation because the use of co-ocurrence frequencies. The calculus is showed by the next formula where "i" and "j" are adjacent image points values under one pre-defined direction. p(i,j) is the probability of that co-ocurrence over the image.
+{{ interimage:att_stddeviationglcm.gif }}
 ===== Neighborhood Attributes =====
@@ Linha 106: / Linha 159: @@
   * **meanDifferenceTo** - mean difference of the layer **L** mean value of an image object to the layer **L** mean value of all image objects belonging to the selected class **C**.
-  * **meanDifferenceToNeighbors** - mean difference of the layer **L** mean value of an image object to the layer **L** mean value of direct neighbors (**R**=0) or all neighbor objects inside the range **R** (**R**>0). The differences are weighted with regard to the shared border (**R**=0) or the area covered by the neighbor objects inside the range (**R**>0). The distance between the image object and its neighbors is calculated considering their centroids.
+  * **meanDiffToNeighbors** - mean difference of the layer **L** mean value of an image object to the layer **L** mean value of direct neighbors (**R**=0) or all neighbor objects inside the range **R** (**R**>0). The differences are weighted with regard to the shared border (**R**=0) or the area covered by the neighbor objects inside the range (**R**>0). The distance between the image object and its neighbors is calculated considering their centroids.