DETEKCIJA RAZPOK V STEKLU Z METODAMI STROJNEGA VIDA

• Main Author: IVANOVSKA, MARIJA
• Other Authors: Perš, Janez
• Format: Master Thesis
• Language: Slovenian
• Published: 2019
• Subjects: strojni vid; razpoke v steklu; Gaborjevi filtri; SVM; globoko učenje; CNN; VGG16; machine vision; glass cracks; Gabor filters; deep learning; Osnova; sami
• Online Access: https://repozitorij.uni-lj.si/IzpisGradiva.php?id=106001; https://repozitorij.uni-lj.si/Dokument.php?id=116700&dn=

Raziskava se osredotoča na razvoj algoritmov strojne inteligence z uporabo konvencionalnih metod ter z uporabo nevronskih mrež. Izziv, s katerim smo se pri razvoju metod ukvarjali, se nanaša na iskanje poškodb na steklu vial. Osnova za razvoj algoritmov je zbirka slik poškodovanih in nepoškodovanih vial, ki smo jo ustvarili sami s primerno izbrano opremo in ustreznimi postopki. Prvi predlagani pristop iskanja poškodb zajema klasične metode strojnega učenja. Za luščenje značilk smo uporabili banko Gaborjevih filtrov, za razvrščevanje vzorcev pa je bila uporabljena metoda podpornih vektorjev. Drugi predlagani pristop zajema metode globokega učenja. V tem primeru smo kot osnovni model za konvolucijo vzorcev s filtri vzeli arhitekturo mreže VGG16, ki ji je bil odstranjen zgornji del. Namesto tega smo implementirali nove plasti s prilagojenim številom parametrov in previdno izbranimi aktivacijskimi funkcijami. V obeh primerih smo za evaluacijo algoritmov uporabili ROC krivulje, pri čemer smo dosegli 100% razpoznavnost poškodb. Zaradi optimizacije časa izvajanja teh algoritmov smo na koncu izvedli še ablacijsko študijo, pri kateri smo opazovali, kako zmanjševanje podatkov o eni viali vpliva na končni izzid razvrščanja. Focus of this thesis is on development of artificial intelligence algorithms in two ways, first by using standard machine vision methods and then by using neural networks. The main problem we were solving was detection of glass cracks on vials. For this purpose, with carefully chosen procedures and equipment we created a dataset of damaged and undamaged vials. First proposed algorithm includes classic machine vision methods. Extraction of image features was done with the help of a bank of Gabor filters while sorting of vials was done using trained support vector machines. Second proposed algorithm is actually a deep learning method. It consists of the convolutional neural network VGG16, without its fully connected layers on the top. Instead of them, some fully connected layers with adapted quantity of parameters were added. Both algorithms were evaluated using ROC curves and they both gained 100% accuracy in recognizing damaged as well as undamaged vials. To optimize both algorithms in term of time needed for processing data, we also did an ablation study where we were systematically removing features from the model to see how relevant they are for the final result.