Endüstriyel uygulamalarda görüntü işleme tabanlı otomatik hata tespit yöntemlerinin geliştirilmesi
Yükleniyor...
Dosyalar
Tarih
2021
Yazarlar
Dergi Başlığı
Dergi ISSN
Cilt Başlığı
Yayıncı
BTÜ, Fen Bilimleri Enstitüsü
Erişim Hakkı
info:eu-repo/semantics/openAccess
Özet
Günümüzde maliyet, hız ve verimliliği en iyi noktaya getirmek adına insan faktörün aradan çıkarılarak bilgisayar tabanlı üretim süreçleri kullanılmaktadır. Gerçek zamanlı ve kameralı kontrol sistemleri sayesinde, yapılmak istenilen kalite kontrol süreçleri daha doğru ve güvenilir bir şekilde yapılabilmektedir. Bu çalışmada, görsel kalite kontrol sürecini daha hızlı ve verimli hale getirmek için iki aşamalı bir otomatik hata tespit sistemi geliştirilmiştir. Tezin ilk bölümünde, kameraya bağlı sensör, lens ve aydınlatma seçimleri yapılarak görüntünün daha doğru şekilde algılanması süreci ele alınmıştır. Sensör kamera için en önemli etken olduğundan, çalışma mesafesine bağlı olarak yakalanabilecek minimum hata boyutu göz önüne alınarak sensör seçimi yapılmıştır. Lens seçimi sensöre, çalışma mesafesi ve uygulanan görüş alanı değerince çeşitli parametreler hesaplanarak yapılmıştır. Aydınlatma seçimi ise, nesne yüzeyi ve çalışma mesafesine bağlı olarak yapılmıştır. Seçimlerden sonra, hata tespit sistemi için en ideal haberleşme arayüzü belirlenmiş ve test düzeneğinin mekanik aksamları hazırlanmıştır. Tezin ikinci bölümünde ise hataların otomatik tespiti için kullanılacak kalite kontrol süreci ve uygun görüntü işleme yöntemleri incelenmiştir. Uygun yöntemlerin belirlenmesinden önce, algılanan nesnenin RGB (Red, Green, Blue)'den HSL (Hue, Saturation, Luminance)'e dönüştürme işlemi yapılmıştır. Çalışmada, endüstriyel olarak kullanılabilecek otomatik bir hata kontrol sisteminin geliştirilmesi ve uygun görüntü işleme tekniklerinin belirlenmesi hedeflenmiştir. Kalite kontrol süreci, test edilen örnek parçadaki bir desen eşleştirme tekniği olan logo tanımlama üzerine kurulmuştur. LabVIEW ve ilgili görüntü işleme araçları kullanılarak bu yöntemler gerçek zamanlı bir otomatik hata tespit yazılımı olarak gerçekleştirilmiştir. Geliştirilen yazılım ile endüstriyel uygulamalarda sıkça karşılaşılan çap kontrolü, var/yok kontrolü, optik karakter tanıma, barkod okuma, desen kontrolü ve kenar kontrolü gibi farklı hatalar otomatik olarak tespit edilebilmektedir. Daha sonra sistem için uygunluğu test edilen desen eşleştirme tekniklerinden, piramit eşleştirme ve düşük tutarsızlık örnekleme algoritmaları aynı doğruluk değerleri altında zaman açısından karşılaştırılmıştır. Bu işlemden sonra, üretim sürecinde oluşan çizik gibi fiziksel kusurların tespiti ve miktarı histogram eşleştirme normalizasyonu ile yapılmıştır. Yüzey deformasyon tespitinden sonra, model 2 tipi seçilen QR (Quick Response) kodu okunmuştur. Daha sonra, görüntü üzerinde yer alan deliklerin algılanması geometrik eşleştirme ve çaplarının ölçümü kenar tabanlı algoritma kullanılarak yapılmıştır. Deliklerin algılanması ve çap ölçümünden sonra, Optik Karakter Tanıma (OCR - Optical Character Recognition) yapılmıştır. Optik karakter tanıma işlemi için bir eğitim kümesi oluşturulup, eğitim kümesine verilerine dayalı olarak oto lineer modda tanıma işlemi yapılmıştır. Son olarak kenar tespit işlemi, basit kenar tespit yöntemi ve Hough dönüşümü kullanılarak yapılmıştır. İşlenmiş bir metal parçası üzerinde elde edilen veriler test sonuçlarında verilmiştir. Gerçekleştirilen testlerde, kullanılan yöntemlerde eşik değer belirleme probleminin hatanın tespitinde problemlere yol açtığı görülmüştür. Bu nedenle, testlerden elde edilen verilerin yapay sinir ağı (YSA) kullanılarak, eşik değerlerin adaptif olarak belirlenebileceği "YSA Tabanlı Eşik Değer Belirleme Yöntemi" önerilmiştir. Zaman açısından daha iyi sonuç veren algoritmanın en etkili parametreleri, akıllı bir yöntem olan YSA kullanılarak optimize edilmiştir. Eşik değerlerin belirlenmesinde Levenberg–Marquardt (LM), ölçekli eşlenik gradyen (scaled conjugate gradient-SCG) ve Bayesian düzenlemesi (Bayesian regularization-BR) algoritmaları karşılaştırılmış ve en optimum değeri sağlayan algoritma seçilmiştir. Düşük tutarsızlık örnekleme en etkin parametreleri Matlab kullanılarak YSA ile eğitilmiştir. Elde edilen sonuçlara göre, önerilen yöntemin gerçek zamanlı uygulamalar için kararlı sonuçlar verdiği görülmüştür.
Nowadays, computer-based production processes are used by removing the human factor for bring the cost, speed and efficiency to the best point. With real-time and camera control systems, quality control processes required can be done more accurately and reliably. In this study, a two-stage automatic error detection system has been developed to make the visual quality control process faster and efficient. In the first part of the thesis, the process of perceiving the image more accurately by making the camera-dependent sensor, lens and lighting selections is discussed. Wherefore the sensor is the most important factor for the camera, the sensor has been selected considering the minimum error size that can be caught depending on the working distance. Lens selection was made by calculating various parameters according to the sensor, working distance and applied field of view. Lighting selection was made depending on the object surface and working distance. After the selections, the most ideal communication interface for the fault detection system was determined and the mechanical parts of the test setup were prepared. In the second part of the thesis, the quality control process and appropriate image processing methods to be used for automatic detection of errors are examined. Before determining the appropriate methods, the detected object was transformed from RGB (Red, Green, Blue) to HSL (Hue, Saturation, Luminance). In the study, it is aimed to develop an automatic error control system that can be used industrially and to determine appropriate image processing techniques. The quality control process is based on logo identification, a pattern matching technique on the sample piece tested. Using LabVIEW and related image processing tools, these methods were implemented as a real-time automatic error detection software. With the developed software, different errors such as diameter control, pass-fail control, optical character recognition, barcode reading, pattern control and edge control, which are frequently encountered in industrial applications, can be detected automatically. Then, pyramid matching and low inconsistency sampling algorithms, among the pattern matching techniques tested for the system, were compared in terms of time under the same accuracy values. After this process, the detection and amount of physical defects such as scratches in the production process were made by histogram matching normalization. After detecting the surface deformation, the QR (Quick Response) code of the model 2 type was read. Then, detection of holes on the image, geometric matching and measurement of their diameters were made using an edge-based algorithm. Optical Character Recognition (OCR) was performed after detecting the holes and measuring the diameter. A training set was created for the optical character recognition process, and the recognition process was performed in auto linear mode based on the training set data. Finally, edge detection was done using the simple edge detection method and Hough transform. The data obtained on a machined piece of metal are given in the test results. In the tests carried out, it was observed that the problem of determining the threshold value in the methods used caused problems in the detection of the error. For this reason, "ANN-Based Threshold Value Determination Method" has been proposed, in which the threshold values can be determined adaptively by using the data obtained from the tests using an artificial neural network (ANN). The most effective parameters of the algorithm, which gives better results in terms of time, have been optimized by using an intelligent method, ANN. Levenberg-Marquardt (LM), scaled conjugate gradient (SCG) and Bayesian regularization (BR) algorithms were compared in determining the threshold values and the algorithm providing the optimum value was selected. The most efficient parameters of low inconsistency sampling were trained with ANN using Matlab. According to the results, it is seen that the proposed method gives stable results for real time applications.
Nowadays, computer-based production processes are used by removing the human factor for bring the cost, speed and efficiency to the best point. With real-time and camera control systems, quality control processes required can be done more accurately and reliably. In this study, a two-stage automatic error detection system has been developed to make the visual quality control process faster and efficient. In the first part of the thesis, the process of perceiving the image more accurately by making the camera-dependent sensor, lens and lighting selections is discussed. Wherefore the sensor is the most important factor for the camera, the sensor has been selected considering the minimum error size that can be caught depending on the working distance. Lens selection was made by calculating various parameters according to the sensor, working distance and applied field of view. Lighting selection was made depending on the object surface and working distance. After the selections, the most ideal communication interface for the fault detection system was determined and the mechanical parts of the test setup were prepared. In the second part of the thesis, the quality control process and appropriate image processing methods to be used for automatic detection of errors are examined. Before determining the appropriate methods, the detected object was transformed from RGB (Red, Green, Blue) to HSL (Hue, Saturation, Luminance). In the study, it is aimed to develop an automatic error control system that can be used industrially and to determine appropriate image processing techniques. The quality control process is based on logo identification, a pattern matching technique on the sample piece tested. Using LabVIEW and related image processing tools, these methods were implemented as a real-time automatic error detection software. With the developed software, different errors such as diameter control, pass-fail control, optical character recognition, barcode reading, pattern control and edge control, which are frequently encountered in industrial applications, can be detected automatically. Then, pyramid matching and low inconsistency sampling algorithms, among the pattern matching techniques tested for the system, were compared in terms of time under the same accuracy values. After this process, the detection and amount of physical defects such as scratches in the production process were made by histogram matching normalization. After detecting the surface deformation, the QR (Quick Response) code of the model 2 type was read. Then, detection of holes on the image, geometric matching and measurement of their diameters were made using an edge-based algorithm. Optical Character Recognition (OCR) was performed after detecting the holes and measuring the diameter. A training set was created for the optical character recognition process, and the recognition process was performed in auto linear mode based on the training set data. Finally, edge detection was done using the simple edge detection method and Hough transform. The data obtained on a machined piece of metal are given in the test results. In the tests carried out, it was observed that the problem of determining the threshold value in the methods used caused problems in the detection of the error. For this reason, "ANN-Based Threshold Value Determination Method" has been proposed, in which the threshold values can be determined adaptively by using the data obtained from the tests using an artificial neural network (ANN). The most effective parameters of the algorithm, which gives better results in terms of time, have been optimized by using an intelligent method, ANN. Levenberg-Marquardt (LM), scaled conjugate gradient (SCG) and Bayesian regularization (BR) algorithms were compared in determining the threshold values and the algorithm providing the optimum value was selected. The most efficient parameters of low inconsistency sampling were trained with ANN using Matlab. According to the results, it is seen that the proposed method gives stable results for real time applications.
Açıklama
Anahtar Kelimeler
Elektrik ve Elektronik Mühendisliği, Electrical and Electronics Engineering, Sayısal görüntü işleme, Digital image processing