Search

V nalogi je predstavljen sistem računalniškega vida, ki temelji na obdelavi slike in metodah strojnega učenja, in je namenjen avtomatski oceni kalivosti semen paradižnika (Lycopersicon lycopersicum L.). Celoten sistem je bil zgrajen z uporabo odprtokodnih aplikacij ImageJ, Weke in njunih javnih javanskih razredov ter povezan s pomočjo namensko razvite programske kode. V raziskavi nismo uporabljali komercialne programske opreme. S pomočjo osmih algoritmov strojnega učenja to so naivni Bayesov klasifikator (NBC), k-najbližjih sosedov (k-NN), odločitvena drevesa, metoda podpornih vektorjev (SVM), umetna nevronska mreža, bagging, boosting in naključni gozdovi, smo izgradili klasifikacijske modele in jih medsebojno primerjali na vzorcu 700 kalečih semen. Najboljšo povprečno klasifikacijsko točnost 95,743 % s standardnim odklonom 2,56 ob desetkrat ponovljenem desetkratnem prečnem preverjanju je dosegel model umetne nevronske mreže (večplastni perceptron). Izboljšan t-test (resampled t-test) s stopnjo zaupanja 0,05 je pokazal, da se rezultat statistično značilno razlikuje od preostalih testiranih modelov klasifikatorjev. Sledili so model SVM (94,114 % ± 2,60), bagging (94,071 % ± 2,84), naključni gozdovi (93,743 % ± 2,93), k-NN (93,714 % ± 2,42), odločitvena drevesa (93,586 % ± 2,72) in model boosting (93,443 % ± 3,01), vendar razlike med njimi niso statistično značilne. Najnižjo povprečno klasifikacijsko točnost je dosegel model NBC (87,929 % ± 4,09), katerega razlika je bila statistično značilna. Ker je model umetne nevronske mreže dosegel najboljše rezultate tudi pri preciznosti, priklicu, F-meri in ploščini pod krivuljo ROC, smo ga uporabili za izgradnjo prototipa, namenjenega klasifikaciji semen paradižnika, kalečih v petrijevkah (90 x 98 x 18mm). Prototip je pravilno klasificiral več kot 95 % kalečih semen. This thesis describes a computer vision system, based on image processing and machine learning techniques, which was implemented for automatic assessment of germination rate of the tomato seeds (Lycopersicon lycopersicum L.). The entire system was built using the open source applications ImageJ, Weka and their public Java classes and was linked by a specially developed code. No expensive commercial software was used. Eight machine learning classification algorithms, Naive Bayes classifiers (NBC), k-nearest neighbours (k-NN), decision trees (DT), support vector machines (SVM), artificial neural networks (ANN), bagging, boosting and random forests were implemented and directly compared on a sample of 700 seeds for the first time. A 10-fold cross-validation repeated 10 times was used in evaluating the performance of all classification models. The best average classification accuracy of 95,743 % with standard deviation of 2,56 was obtained with the artificial neural networks (multilayer perceptron), which is significantly more accurate than SVM (94.114% ± 2.60), bagging (94.071% ± 2.84), random forests (93.743% ± 2.93), k-NN (93.714 ± 2.42%), decision trees (93.586 % ± 2.72), and boosting (93.443% ± 3.01). The lowest average classification accuracy was reached by NBC model (87.929% ± 4.09). Since the model ANN also achieved the best results for the precision, recall, F-measure, and the area under the ROC curve, we incorporated it in a prototype intended for the classification of germinated tomato seeds. The automated system was able to correctly classify 95 % of germinated tomato seeds in Petri dishes (90x98x18mm).