The databases available here for download use the ARFF format, used for Weka data mining software.

I used the databases for my master thesis project. Each database contains image descriptors (concerning the shape, color or texture), computed by a feature extraction algorithm programmed with Matlab.

The images were either produced by me and my colleges at the LIVIA, either downloaded from the internet (to see the source, follow the link on the Reference column of the following table).

The table shows information about the 8 arff files available for download. Use this link to download all 8 files at once.

File	Nb. classes	Nb. ins./clas.	Total ins.	Reference	Description	Best median of error obtained
cereals_500.arff	6	500	3000	* See below	6 types of cereals, segmented	SVM2 : 1.5% error
cereals_950.arff	6	950	5700	* See below	same than preceding but with more instances per class	SVM2 1.5% error
leaves_60.arff	3	60	180	Caltech	3 types of leaves, no segmentation	SVM2 : 26.67% error
cropped_leaves_60.arff	3	60	180	* See below	3 types of leaves - cropped images (rectangles)	GPB, SVM or MP (equal) : 6.67% error
digits_33.arff	10	60	330	* See below	Computer characters, gray-scale, segmented	MP : 5.45% error
knots_27.arff	6	27	162	Oulu	Wood defects, segmented	MP : 20.37% error
pollen_196.arff	7	196	1372	Bangor	7 types of pollen, gray-scale, segmented	SVM2 : 4.15% error
raisins_450.arff	3	450	1350	* See below	3 types of raisins, segmented	SVM : 0.22% error

* The images that were used to produce theses databases were made by students and researchers of the LIVIA. The image files are not available here because of their size. If you would like to use them, don't hesitate to contact me.

The databases of the table contain 60 features for RGB images, 48 for the gray-scale images and 43 for the RGB images without segmentation (leaves only). The last column named Best median of error obtained, shows the median of classification error, on 50 experiments with 2/3 training and 1/3 test random splitting of the data, for the best classifier from the following list (all from Weka):

• Bayes : Naive Bayes classifier
  
• C4.5 : C4.5 Decision Tree. Called J48 in Weka
  
• KNN : K Nearest Neighbors, with K=3. Called IBk in Weka
  
• GP : Genetic Programming without boosting (earlier version of GPB, see below)
  
• SVM : Support Vector Machine, with 1st degree polynomial function. SMO in Weka
  
• SVM2 : Support Vector Machine, 2nd degree polynomial function
  
• MP : Multilayer Perceptron
  
• Boosted versions of Bayes, C4.5, SVM, SVM2 et MP, with AdaBoost: when boosted, the classifier identifiers are precede by A, (ex: ABayes, AMP, etc.)
  
• GPB : Genetic Programming with integrated boosting, developed for Weka

For questions or comments, please contact me.