• No results found

The final stage is the classification stage. The resulting images after all the processing are finally examined for the presence of any type of defect out of the four defects. The defective image is then classified to a particular defect type.

Different classification procedures have been used for different defects. Template matching has been tested on various defects for classification. Neural networks are the next plan to perform the classification. The existing neural network functions have been applied and tested. To date, more work need to be undertaken to complete this work. The future work might be based on neural networks depending on the quality of results achieved in near future. Currently they are under consideration and being researched on.

4.12.1 Template Matching

Template matching is a technique in image processing where an image is segmented or its areas of interests are separated out from the image using a small template. It may also be defined as comparing the parts of the image against each other. The template or the mask or the sample image is matched by moving it to all possible locations on the source image and where ever it finds something similar, it is marked out. The result is numerical index of the highly similar locations.

Template matching can either be feature based or template based. Feature based method uses features such as corners or edges to do the search. Template approach uses the entire template to perform the matching. Matching is performed pixel by pixel and fast template matching can be performed by using cross correlation method (Brunelli 2009) (Lewis 1995).

Inbuilt template matching function of Matlab was tested on wire defects to get the results shown in Figure 4.21. The result image clearly shows extra noise and non- defect detection at the same time. Hence, template matching technique isn‟t of much success for the purpose.

65

Figure 4.21: Results of wire defect detection using template matching

4.12.2 Artificial Neural Networks

Neural networks are an advanced technique of computer programming. They are efficient at pattern recognition and solving other such problems. An Artificial neural network works in the same way as biological nervous system (brain) works. It is made up of large number of interconnected units, called neurons that are capable of processing information. These neurons work in coordination with each other to solve some problem. As human brain learns by experiences or examples, artificial networks are also configured for a specific application through a learning process. A neuron is the basic processing unit of a neural network. They receive input in the form of information from some source, process them according by using the learned behaviours, and output the results.

Input layer neurons get input from some real source and the output layer provides the output. All the other processing neurons are hidden. Single neuron is not really helpful, but their combination into multilayer structures, called neural networks, is a solution to many problems.

Neural network has been used in combination to edge detection algorithms and morphological operators in the research. The detection results produced by neural networks have been presented and discussed in the next chapter.

The Feed Forward Back-Propagation architecture is currently the most efficient, simple and popular neural network solution for complex, multi-layered networks. It is used in many applications, as provides workable non-linear solutions to ill-defined problems.

66

The typical model has an input layer, at least one or more hidden layer(s) and an output layer. Theoretically, there is no limit on the number of hidden layers but typically they are one or two. Feedback is the basic characteristic stage of this type of network. In this stage the output of one layer is directed back to a previous layer. The information flows through in and out layers during recall. The process of simulation of a network with an unknown set of data to get the answer is called recall. Back- propagation occurs only during training and is not used during recall.

4.13 Summary

This chapter has outlined all the image processing techniques that have been tested in order to develop an efficient detection method for the defects under investigation in this research. It has also covered the methods that have been used as part detection in the final developed algorithms. The chapter has been sectioned into pre- processing, processing, post processing and the classification stages for the better understanding of the imaging processes. Next chapter explains the defect detection algorithms developed for the current research in detail.

67

5 Developed Defect Detection Algorithms

5.1 Introduction

This chapter presents the description of the algorithms developed specifically to process the JLI images together with some of the results of the defect detection being produced on individual defect images provided by Tata Steel, for each type of defect. The offline images were to be treated by the software and a report of the detection was shown on screen for experts to assess.

The software was run on an Intel Core I7 CPU 920 at 2.67 GHz, with 8 GB of RAM, Windows XP 64 Bits Professional. The software was developed using the OpenCV image processing library encapsulated in EMGU for C# (Microsoft Visual Studio Professional). The running time for a rail of on average 110 m long with on average 440 images per camera view is on average 55 s. The time between each rail while in heavy production is more than 1 min 30 seconds, making the software capable of analysing rails in real-time.

68