magistrsko delo
Marcel Petek (Author), Dušan Gleich (Mentor), Simon Klančnik (Mentor)

Abstract

Magistrsko delo podaja pregled metod za propoznavo površinskih napak na obdelovancih. Objekti opazovanja so krmilne tipke, podsklop ohišij v elektroomaricah. Predstavljene bodo tri metode razvrščanja krmilnih tipk v dober in slab razred. Zajemanje slik je bilo opravljeno s pomočjo laboratorijske opreme, saj so elementi opreme višjega cenovnega razreda. Namen magistrske naloge je v bazah slik krmilnih tipk z različnimi metodami prepoznati napake in jih razvrstiti v pripadajoči razred. Zaradi specifičnosti problematike zaznavanja so se metode prilagajale problemu. Uporabili smo metode prepoznave napak na nadzorovan in nenadzorovan način, torej globinsko učenje z uporabo nevronske mreže, avtoenkoderja in klasično pragovno metodo z uporabo različnih detektorjev robov in preglednih tabel. Omenjene globoke metode se dandanes ne uporabljajo v veliki meri za industrijske namene. Metode so se namreč izboljšale do te mere, da veliki koncerni, kot so IBM, Google, Facebook, uporabniku napram preteklim iskalnim nizom v brskalniku predlagajo, kaj naj bi iskal po svetovnem spletu. Za izbiro globokega učenja namesto genetskega ali algoritma rojev delcev smo se odločili izključno zaradi hitre prilagoditve programa na vhodne parametre in razvoja programa od preteklosti, ko je nivo globine nevronskih mrež bila samo ena prikrita plast z enim nevronom, do danes, ko se lahko nivo adaptivno spreminja glede na vhodno problematiko. Dostopni algoritmi za zaznavanje defektov na teksturah, ki smo jih preizkusili v komercialnih paketih (Vision NI), niso bili učinkoviti za detekcijo teh nepravilnosti. To je motivacija za raziskovanje učinkovitosti drugih pristopov in za primerjavo učinkovitosti. S primerjavo metod bomo za nadaljnje raziskovanje izbrali tisto, ki bo dosegla cilj, 95-odstotno stopnjo natančnosti razvrstitve v razreda dober in slab. Začetni cilj razvrstitve smo uspeli dosečti z uporabo globokega učenja nevronskih mrež.

Keywords

avtoenkoderji;strojni vid;globoko učenje;nevronske mreže;magistrske naloge;

Data

Language: Slovenian
Year of publishing:
Typology: 2.09 - Master's Thesis
Organization: UM FERI - Faculty of Electrical Engineering and Computer Science
Publisher: M. Petek
UDC: 004.93:620.191(043.2)
COBISS: 22168086 Link will open in a new window
Views: 1138
Downloads: 137
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Machine vision system for surface inspection
Secondary abstract: The master's thesis provides an overview of methods for recognizing surface errors on work pieces. The observation objects are control keys, a sub-assembly of the enclosures in the electric chambers. Three methods for classifying control keys to a good and a bad class will be presented. Image capture was done with the help of laboratory equipment, since the elements of the equipment are of higher price range. The purpose of the master's thesis is to identify faults of control keys with different methods and to classify them in the corresponding class. Due to the specificity of the problem of identification, the methods were adjusted to the problem. We used methods of recognition of errors in a controlled and uncontrolled way, therefore, deep learning with usage of neural networks, auto encoder and classical sluice-gate method, using various detectors of edges and lookup tables. These deep methods are not widely used today for industrial purposes. The methods have improved to such an extent that the great concerns, such as IBM, Google, Facebook, suggest to the user what they should search for on the Internet based on the previous searches. We have chosen deep learning instead of genetic or swarm particles algorithm, exclusively due to the programme’s rapid adjustments to the input parameters and the development of program since the past. The level of depth of neural networks used to be only one disguised with a single layer neuron, and now the level can adaptively change depending on the input problems. Accessible algorithms for detecting defects on textures that we tested in commercial packages (Vision NI) were not effective for detecting these malfunctions. This is the motivation for exploring the effectiveness of other approaches and to compare their efficiency. With comparison of methods, we will choose the one who will achieve the goal of 95% degree of accuracy of the classification in class good or bad for further exploration. The initial aim of the ranking we managed to achieve with the use of deep learning of neural networks.  
Secondary keywords: autoencoders;machine vision;classification;deep learning;neural network;
URN: URN:SI:UM:
Type (COBISS): Master's thesis/paper
Thesis comment: Univ. v Mariboru, Fak. za elektrotehniko, računalništvo in informatiko, Mehatronika
Pages: XIII, 89 str.
ID: 10993060