1. Identification of MPEG-4 FDP Patterns in Human Faces using Data-Mining Techniques Work subsidized by projects: HUMODAN IST-2001-32202 CICYT TIC2001-0931 TIC2002-10743-E Britos, P. pbritos@itba.edu.ar Perales López, F. paco.perales@uib.es Abásolo, M.J. mjabasolo@uib.es García Martínez, R. rgm@itba.edu.ar MPEG-4 is an ISO/IEC standard which defines 84 feature points called Face Definition Parameters (FDPs) to parameterise a face. FDPs are used to personalize a generic face model to a particular face. MPEG-4 FDPs 3.2 3.6 3.4 3.83.11 The main purpose of this work is to induce rules that describe patterns in human faces, that means relations between different dimensions of a face. C5.0 vs. C4.5 More precise rules with C5.0 than with C4.5. More complex rules with C5.0 (left part of the rule has a conjunction) than with C4.5. Cluster vs. entire database More precise rules by analysing the main cluster instead of the whole database. Example: objective field discretized“FH” Rules obtained with C4.5 IF FW >=5 THEN FH range = 5 IF LE >= 84mm THEN FH range = 4 IF LEH < 40mm THEN FH range = 2 IF NH >= 54.6mm THEN FH range = 1 IF weight < 50 kg. THEN range = 1 IF LID >= 25mm THEN FH range = 4 Rules obtained with C5.0 IF NH <= 53.9mm THEN FH range = 1 IF LEH 71mm AND MH <=21mm THEN FH range = 2 IF LEH 23.9mm THEN FH range = 2 IF LEH > 55mm AND LE <= 88mm AND LID < 24mm THEN FH range = 3 IF LEH 71mm AND MH >21mm THEN FH range = 3 IF LEH > 55mm THEN FH range = 4 Discretization of the continuous fields allows using it as an objective of the rules. FIELD DISCRETIZATION SOM are used to classify high- dimensional data. In this work we use SOM for clustering the records. SELF ORGANIZING MAPS DATABASE OF FACES FH: 11.1 to 2.1 FW : 10.10 to 10.9 In our work a face is described by distances between different MPEG-4 FDPs. (i.e. mouth width, eyebrown width, etc.). We have a database of 600 faces of different sex, race, etc. RE: 4.2 to 4.6 REW: 3.12 to 3.8 REH: 3.14 to 3.10 NH: 9.6 to 9.2 NT: 9.3 to 9.15 MW: 8.4 to 8.3 MH: 8.1 to 8.2 RULES FOR WHAT? To personalize a generic face model with standard measures according some conditions like sex, race, etc. To discover relations between different parts of a face To discover relations between the parts of a face and other characteristics like sex, race, height, etc. To classify an unknown face example (sex, race, etc.) Example: objective field “sex” Rules obtained with C4.5 IF weight < 63 kg. THEN sex = female IF NA >= 81,57º THEN sex = female IF weight >= 72 kg. THEN sex = male IF weight < 72 kg. THEN sex = female IF RID >= 23 mm THEN sex = male Rules obtained with C5.0 IF weight < 70 kg. AND LE <=79mm THEN sex = female IF weight <= 62 kg. THEN sex = female IF weight >= 62 kg. AND LE >79mm THEN sex = male IF weight > 70 kg. THEN sex = male C4.5 C4.5 is an automatic learning algorithm for classifying examples. It obtains decision trees or sets of if-then rules forms. C5.0 is an improvement of C4.5. C5.0 DATA MINING TECHNIQUES Data mining is all about extracting patterns from a warehoused data.