Abstract
The defect detection on manufactures is of utmost importance in the optimization of industrial processes (Garcia 2005). In fact, the industrial inspection of engineering materials and products tends to the detection, localization and classification of flaws as quickly and as accurately as possible in order to improve the production quality. In this field a relevant area is constituted by visual inspection. Nowadays, this task is often carried out by a human expert. Nevertheless, such kind of inspection could reveal time-consuming and suffer of low repeatability because the judgment criteria can differ from operator to operator. Furthermore, visual fatigue or loss of concentration inevitably lead to missed defects (Han, Yue & Yu 1999, Kwak, Ventura & Tofang-Sazi 2000, Y.A. Karayiannis, R. Stojanovic, P. Mitropoulos, C.Koulamas, T. Stouraitis, S. Koubias & G. Papadopoulos 1999, Patil, Biradar & Jadhav 2005). In order to reduce the burden of human testers and improve the detection of faulty products, recently many researchers have been engaged in developing systems in Automated Visual Inspection (AVI) of manufactures (Chang, Lin & Jeng 2005, Lei 2004, Yang, Pang & Yung 2004). These systems reveal easily reliable from technical point of view and mimic the experts in the evaluation process of defects appropriately (Bahlmann, Heidemann & Ritter 1999), even if defect detection in visual inspection can become a hard task. In fact, in industrial processes a large amount of data has to be handled and flaws belong to a great number of classes with dynamic defect populations, because defects could present similar characteristics among different classes and different interclass features (R. Stojanovic, P. Mitropulos, C.Koullamas, Y. Karayiannis, S. Koubias & G. Papadopoulos 2001). Therefore, it is needed that visual inspection systems are able to adapt to dynamic operating conditions. To this purpose soft computing techniques based on the use of Artificial Neural Networks (ANNs) have already been proposed in several different areas of industrial production. In fact, neural networks are often exploited for their ability to recognize a wide spread of different defects (Kumar 2003, Chang, Lin & Jeng 2005, Garcia 2005, Graham, Maas, Donaldson & Carr 2004, Acciani, Brunetti & Fornarelli 2006). Although adequate in many instances, in other cases Neural Networks cannot represent the most suitable solution. In fact, the design of ANNs often requires the extraction of parameters and features, during a preprocessing stage, from a suitable data set, in which the most possible defects are recognized (Bahlmann, Heidemann & Ritter 1999, Karras 2003, Rimac-Drlje, Keller & Hocenski 2005). Therefore, methods based on neural networks could be time expensive for inline applications because such preliminary steps and could reveal complex (Kumar 2003, Kwak, Ventura & Tofang-Sazi 2000, Patil, Biradar & Jadhav 2005, R. Stojanovic, P. Mitropulos, C.Koullamas, Y. Karayiannis, S. Koubias & G .Papadopoulos 2001). For this reason, when in an industrial process time constraints play an important role, a hardware solution of the abovementioned methods can be proposed (R. Stojanovic, P. Mitropulos, C.Koullamas, Y. Karayiannis, S. Koubias & G .Papadopoulos 2001), but such kind of solution implies a further design effort which can be avoided by considering Cellular Neural Networks (CNNs) (Chua & Roska 2002). Cellular Neural Networks have good potentiality to overcome this problem, in fact their hardware implementation and massive parallelism can satisfy urgent time constrains of some industrial processes, allowing the inclusion of the diagnosis inside the production process. In this way the defect detection method could enable to work in real time according to the specific industrial process.
TopIntroduction
The defect detection on manufactures is of utmost importance in the optimization of industrial processes (Garcia 2005). In fact, the industrial inspection of engineering materials and products tends to the detection, localization and classification of flaws as quickly and as accurately as possible in order to improve the production quality. In this field a relevant area is constituted by visual inspection. Nowadays, this task is often carried out by a human expert. Nevertheless, such kind of inspection could reveal time-consuming and suffer of low repeatability because the judgment criteria can differ from operator to operator. Furthermore, visual fatigue or loss of concentration inevitably lead to missed defects (Han, Yue & Yu 1999, Kwak, Ventura & Tofang-Sazi 2000, Y.A. Karayiannis, R. Stojanovic, P. Mitropoulos, C.Koulamas, T. Stouraitis, S. Koubias & G. Papadopoulos 1999, Patil, Biradar & Jadhav 2005).
In order to reduce the burden of human testers and improve the detection of faulty products, recently many researchers have been engaged in developing systems in Automated Visual Inspection (AVI) of manufactures (Chang, Lin & Jeng 2005, Lei 2004, Yang, Pang & Yung 2004). These systems reveal easily reliable from technical point of view and mimic the experts in the evaluation process of defects appropriately (Bahlmann, Heidemann & Ritter 1999), even if defect detection in visual inspection can become a hard task. In fact, in industrial processes a large amount of data has to be handled and flaws belong to a great number of classes with dynamic defect populations, because defects could present similar characteristics among different classes and different interclass features (R. Stojanovic, P. Mitropulos, C.Koullamas, Y. Karayiannis, S. Koubias & G. Papadopoulos 2001). Therefore, it is needed that visual inspection systems are able to adapt to dynamic operating conditions. To this purpose soft computing techniques based on the use of Artificial Neural Networks (ANNs) have already been proposed in several different areas of industrial production. In fact, neural networks are often exploited for their ability to recognize a wide spread of different defects (Kumar 2003, Chang, Lin & Jeng 2005, Garcia 2005, Graham, Maas, Donaldson & Carr 2004, Acciani, Brunetti & Fornarelli 2006). Although adequate in many instances, in other cases Neural Networks cannot represent the most suitable solution. In fact, the design of ANNs often requires the extraction of parameters and features, during a preprocessing stage, from a suitable data set, in which the most possible defects are recognized (Bahlmann, Heidemann & Ritter 1999, Karras 2003, Rimac-Drlje, Keller & Hocenski 2005). Therefore, methods based on neural networks could be time expensive for in-line applications because such preliminary steps and could reveal complex (Kumar 2003, Kwak, Ventura & Tofang-Sazi 2000, Patil, Biradar & Jadhav 2005, R. Stojanovic, P. Mitropulos, C.Koullamas, Y. Karayiannis, S. Koubias & G .Papadopoulos 2001). For this reason, when in an industrial process time constraints play an important role, a hardware solution of the abovementioned methods can be proposed (R. Stojanovic, P. Mitropulos, C.Koullamas, Y. Karayiannis, S. Koubias & G .Papadopoulos 2001), but such kind of solution implies a further design effort which can be avoided by considering Cellular Neural Networks (CNNs) (Chua & Roska 2002).
Key Terms in this Chapter
Automated Visual Inspection: An automatic form of quality control normally achieved using one or more cameras connected to a processing unit. Automated Visual Inspection has been applied to a wide range of products. Its target consists of minimizing the effects of visual fatigue of human operators who perform the defect detection in a production line environment.
Region of Interest: A selected subset of samples within a dataset identified for a particular purpose. In image processing, the Region of Interest is identified by the boundaries of an object. The encoding of a Region of Interest can be achieved by basing its choice on: (a) a value that may or may not be outside the normal range of occurring values; (b) purely separated graphic information, like drawing elements; (c) separated semantic information, such as a set of spatial and/or temporal coordinates.
Artificial Neural Networks: A set of basic processing units which communicate to each other by weighted connections. These units give rise a parallel processing with particular properties such as the ability to adapt or learn, to generalise, to cluster or organise data, to approximate non-linear functions. Each unit receives an input from neighbours or external sources and uses it to compute an output signal. Such signal is propagated to other units or is a component of the network output. In order to map an input set into an output one a neural network is trained by teaching patterns, changing its weights according to proper learning rules.
Cellular Neural Networks: A particular circuit architecture which possesses some key features of Artificial Neural Networks. Its processing units are arranged in an M×N grid. The basic unit of Cellular Neural Networks is called cell and contains linear and non linear circuit elements. Each cell is connected only to its neighbour cells. The adjacent cells can interact directly with each other, whereas cells not directly connected together may affect each other indirectly because of the propagation effects of the continuous time dynamics.
Image Matching: Establishment of the correspondence between each pair of visible homologous image points on a given pair of images, aiming at the evaluation of novelties.
Defect Detection: Extraction of information about the presence of an instance in which a requirement is not satisfied in industrial processes. The aim of Defect Detection consists of highlighting manufactures which are incorrect or missing functionality or specifications.
Industrial Inspection: Analysis pursuing the prevention of unsatisfactory industrial products from reaching the customer, particularly in situations where failed manufactures can cause injury or even endanger life.